**2.3 Detecting yellow rust in winter wheat by spectral knowledge base**

In most cases, statistical models for monitoring the disease severity of yellow rust are based on hyperspectral information. The high cost and limited cover of airborne hyperspectral data make it impossible to apply such data for large scale monitoring. Furthermore, the established models of disease detection cannot be used for most satellite images because of the wide range of wavelengths in multispectral images (Zhang et al., 2011).

To resolve this dilemma, the study presents a novel approach by constructing a spectral knowledge base (SKB) of winter wheat diseases, which takes the airborne images as a medium and links the disease severity with band reflectance from moderate resolution remotely sensed data, such as environment and disaster reduction small satellite images (HJ-CCD) accordingly. To achieve this goal, several algorithms and techniques for data conversion and matching are adopted in the proposed system, including minimum noise fraction (MNF) transformation and pixel purity index (PPI) function. The performance of SKB is evaluated with both simulated data and field measured data.

#### **2.3.1 materials and methods**

Experimental design and field conditions was same as the part of 1.1.1.1

#### **2.3.1.1 Inspection of disease severity**

Please refer to the part of 1.2.1.2 above.

#### **2.3.1.2 Airborne hyperspectral imaging**

Please refer to the part of 1.2.1.4 above about airborne hyperspectral imaging and image processing.

#### **2.3.1.3 Acquisition of moderate resolution satellite images**

In this study, the SKB is designed to fit the charge coupled device (CCD) sensor, which is on the environment and disaster reduction small satellites (HJ-1A/B). The basic parameters of the CCD sensor (using 'HJ-CCD' in the following) are given in Table.5. The four bands of


Table 5. Properties of the environment and disaster reduction small satellites (HJ-CCD)

Crop Disease and Pest Monitoring by Remote Sensing 51

shaped areas that were within the field and comprised 7918 'crop-only' pixels for

The reversion model construction was the first step of establishing the SKB. Based on the conclusion of the part above, PRI was a suitable vegetation index for monitoring the severity of yellow rust disease in winter wheat. Therefore, in this study, PRI was used to establish the linkage between the disease severity and the hyperspectral data. Specifically, the yellow rust infection would be apparent at anthesis stage, and this should be closely related with the subsequent yield loss. Therefore, we chose the PHI image at this stage to form the SKB. To obtain a better fitting model, we reanalyzed the PHI-PRI and corresponding DI (%) data at the anthesis stage specifically, and obtained a linear regression model. It should be noted that the data range of DI must be between 0 and 100%. Any predicted DI results that were>100% or <0% were redefined as DI = 100% and DI = 0% to represent very infected

The second step of constructing the SKB is to transform the hyperspectral reflectance of PHIpixels to wide band reflectance of HJ-pixels. To achieve this goal, the best approach is the inherent relative spectral response (RSR) function of the HJ-CCD sensor. By integrating the hyperspectral reflectance of PHI-pixels on the RSR function, the band reflectance of HJ-CCD sensor was thus obtained. Besides, although the wavelength range of the fourth band of HJ-CCD sensor (760 nm-900 nm) was slightly exceeded the maximum wavelength of PHI sensor (850 nm), for most ground measured spectra, the reflectance basically kept on steady from 760 nm to 900 nm. Hence, the simulating results generated using the incomplete range of wavelength (760nm-850nm) should approach to the true value. The integration can be

( )

*end*

*b*

*TM b R f x dx*

**2.3.1.7 Spectral characteristics of different degrees of disease severity** 

image, and to form the typical spectrum for each severity class.

*start*

where *RTM* is the simulated reflectance of a certain band; bstart and bend indicate the beginning and the end wavelength of this band respectively; *f(x)* indicates the RSR function,

Another way to define the disease severity of an undefined pixel, apart from the DI (%) value, is to quantify disease severity by severity classes. The criterion of severity class provided by Huang et al. (2007) was adopted, which corresponded to the major physiological alteration of diseased plants. The DI (%) thresholds for each severity class were: DI<1% indicated not infected (NI), 1%<DI<10% indicated a low degree of infection (LI), 10%<DI<45% indicated mid-range infection (MI), 45%<DI<80% indicated seriously infected (SI) and DI (%)>80% indicated very seriously infected (VI). The MNF transformation and PPI function, which are used for noise reduction and end-member identification, were applied here to select the most representative pixels from the PHI

constructing the SKB. **2.3.1.5 Reversion model** 

shown as follows:

which is obtained from CRESDA.

plants and healthy plants, respectively.

**2.3.1.6 Simulation of the wide band reflectance** 

HJ-CCD covered the visible and near infrared spectral regions. The HJ-CCD sensor has spectral and spatial characteristics that are similar to those of Landsat-5 TM, but the HJ-1A/B satellites have more frequent revisit capability (2 days) than the Landsat-5 satellite (16 days), which is of great importance for agricultural monitoring.

#### **2.3.1.4 Construction of the spectral knowledge base**

The SKB in this study can be interpreted as a pool of relationships between spectral characteristics and prior knowledge. Here, prior knowledge stands for the degree of severity of yellow rust, and the spectral characteristics are the reflectance of the initial four bands of the HJ-CCD image. Hence, there are two major steps involved in constructing the SKB. First, the relationship between hyperspectral information and severity is obtained with a stable empirical reversion model. Then, through the RSR function of the HJ-CCD sensor, the hyperspectal data can be transferred to the wide-band reflectance. In this way, a one-to-one correspondence between the disease severity of yellow rust and reflectances from the HJ-CCD sensor is established at the pixel level. The SKB can represent disease severity in two ways: the DI (%) value and the class of disease severity. The following sections describe each step for establishing the SKB, including data selection, the reversion model, simulation of the wide-band reflectance and estimating the degree of severity. A technical flow diagram of SKB construction is summarized in Fig. 8.

Fig. 8. The flow chart for monitoring of DI(%) of winter wheat stripe rust, b1-b4 represented the reflectance of the four bands of HJ-CCD images

As noted above, the SKB in this study comprised PHI pixels. The predicted accuracy obtained by the SKB was determined primarily by the amount of prior knowledge, which indicated the heterogeneity of disease severity. The design of the yellow rust fungus inoculation ensured a considerable variation in disease severity within the experimental field, from healthy plants to very diseased plants. In addition, to avoid using pixels on or near the ridge in the field that are considered as mixed signals, we chose three rectangular shaped areas that were within the field and comprised 7918 'crop-only' pixels for constructing the SKB.

#### **2.3.1.5 Reversion model**

50 Remote Sensing – Applications

HJ-CCD covered the visible and near infrared spectral regions. The HJ-CCD sensor has spectral and spatial characteristics that are similar to those of Landsat-5 TM, but the HJ-1A/B satellites have more frequent revisit capability (2 days) than the Landsat-5 satellite (16

The SKB in this study can be interpreted as a pool of relationships between spectral characteristics and prior knowledge. Here, prior knowledge stands for the degree of severity of yellow rust, and the spectral characteristics are the reflectance of the initial four bands of the HJ-CCD image. Hence, there are two major steps involved in constructing the SKB. First, the relationship between hyperspectral information and severity is obtained with a stable empirical reversion model. Then, through the RSR function of the HJ-CCD sensor, the hyperspectal data can be transferred to the wide-band reflectance. In this way, a one-to-one correspondence between the disease severity of yellow rust and reflectances from the HJ-CCD sensor is established at the pixel level. The SKB can represent disease severity in two ways: the DI (%) value and the class of disease severity. The following sections describe each step for establishing the SKB, including data selection, the reversion model, simulation of the wide-band reflectance and estimating the degree of severity. A technical flow diagram of

Fig. 8. The flow chart for monitoring of DI(%) of winter wheat stripe rust, b1-b4 represented

As noted above, the SKB in this study comprised PHI pixels. The predicted accuracy obtained by the SKB was determined primarily by the amount of prior knowledge, which indicated the heterogeneity of disease severity. The design of the yellow rust fungus inoculation ensured a considerable variation in disease severity within the experimental field, from healthy plants to very diseased plants. In addition, to avoid using pixels on or near the ridge in the field that are considered as mixed signals, we chose three rectangular

days), which is of great importance for agricultural monitoring.

**2.3.1.4 Construction of the spectral knowledge base** 

SKB construction is summarized in Fig. 8.

the reflectance of the four bands of HJ-CCD images

The reversion model construction was the first step of establishing the SKB. Based on the conclusion of the part above, PRI was a suitable vegetation index for monitoring the severity of yellow rust disease in winter wheat. Therefore, in this study, PRI was used to establish the linkage between the disease severity and the hyperspectral data. Specifically, the yellow rust infection would be apparent at anthesis stage, and this should be closely related with the subsequent yield loss. Therefore, we chose the PHI image at this stage to form the SKB. To obtain a better fitting model, we reanalyzed the PHI-PRI and corresponding DI (%) data at the anthesis stage specifically, and obtained a linear regression model. It should be noted that the data range of DI must be between 0 and 100%. Any predicted DI results that were>100% or <0% were redefined as DI = 100% and DI = 0% to represent very infected plants and healthy plants, respectively.

#### **2.3.1.6 Simulation of the wide band reflectance**

The second step of constructing the SKB is to transform the hyperspectral reflectance of PHIpixels to wide band reflectance of HJ-pixels. To achieve this goal, the best approach is the inherent relative spectral response (RSR) function of the HJ-CCD sensor. By integrating the hyperspectral reflectance of PHI-pixels on the RSR function, the band reflectance of HJ-CCD sensor was thus obtained. Besides, although the wavelength range of the fourth band of HJ-CCD sensor (760 nm-900 nm) was slightly exceeded the maximum wavelength of PHI sensor (850 nm), for most ground measured spectra, the reflectance basically kept on steady from 760 nm to 900 nm. Hence, the simulating results generated using the incomplete range of wavelength (760nm-850nm) should approach to the true value. The integration can be shown as follows:

$$\mathcal{R}\_{TM} = \bigcap\_{b\_{\text{start}}}^{b\_{\text{end}}} f(\mathbf{x})d\mathbf{x}$$

where *RTM* is the simulated reflectance of a certain band; bstart and bend indicate the beginning and the end wavelength of this band respectively; *f(x)* indicates the RSR function, which is obtained from CRESDA.

#### **2.3.1.7 Spectral characteristics of different degrees of disease severity**

Another way to define the disease severity of an undefined pixel, apart from the DI (%) value, is to quantify disease severity by severity classes. The criterion of severity class provided by Huang et al. (2007) was adopted, which corresponded to the major physiological alteration of diseased plants. The DI (%) thresholds for each severity class were: DI<1% indicated not infected (NI), 1%<DI<10% indicated a low degree of infection (LI), 10%<DI<45% indicated mid-range infection (MI), 45%<DI<80% indicated seriously infected (SI) and DI (%)>80% indicated very seriously infected (VI). The MNF transformation and PPI function, which are used for noise reduction and end-member identification, were applied here to select the most representative pixels from the PHI image, and to form the typical spectrum for each severity class.

Crop Disease and Pest Monitoring by Remote Sensing 53

To verify the performance of SKB in identifying and monitoring the severity of yellow rust diseases, two datasets were used: the simulated data and the field-measured data with

The simulated data comprised 50 randomly selected pixels in the same experimental field, but outside the three regions selected for constructing the SKB. The hyperspectral information of each pixel was used to create the reference DI (%) and severity class with the empirical model and the corresponding threshold for each severity class. To test the performance of SKB in terms of DI (%) value, we estimated the DI value with both distance criteria described above. The samples were split into two: the pixels with a reference DI between 1 and 100%, i.e. the 'diseased' pixels, and those with a reference DI<1%, i.e. 'healthy' pixels. For the diseased pixels, the estimated DIs were compared with the reference DI by Pearson correlation analysis and the normalized root mean square error (NRMSE). For the healthy pixels, we used 'yes or no' to determine whether the estimated value indicated infection or not, which also provided an accuracy ratio. The estimation of severity class was verified by overall accuracy and the kappa coefficient.

The field surveyed data sets included the ground investigation of disease severity and the corresponding HJ-CCD images. Between June 1–3, 2009, when the winter wheat was at the anthesis stage, we conducted a survey in the southeast of GanSu Province. The climate of the area surveyed is characterized by high humidity and rainfall, and yellow rust disease occurs almost every year. This area has similar environmental conditions and cultivation customs to those where we constructing the SKB in Beijing, and this makes it an appropriate place for model verification. With the aid of the local Department of Plant Protection, 26 plots were randomly selected and surveyed in the area (Fig. 9). To relate the surveyed value to the pixel value of the HJ-CCD image, we defined the plot as a uniformly planted winter wheat region with an area no less than 30 m in radius. The geographical coordinates of each plot were measured by GPS at the centre of the plot. Disease severity was measured as described above. We repeated the measurement in ve evenlydistributed sections in each plot, and 20 individual plants were included in each measurement. The HJ-CCD images (ID: 122516, 122518) acquired on June 2, 2009 completely covered the surveyed area. The raw data from the HJ-CCD imagery was calibrated based on the corresponding coefficients provided by CRESDA. The calibrated data were atmospherically corrected with the algorithm provided by Liang et al. (2001), which estimated the spatial distribution of atmospheric aerosols and retrieved surface reflectance under general atmospheric and surface conditions. The images were also geometrically corrected against historical reference images with the same geographical coordinates. The images were rectified with a root mean square error of less than 0.5 pixels. The spectrum of the each plot was extracted from the image according to the GPS records. The estimated accuracy in this step followed the same process as the simulated data.

There were 7918 pixels included in the process of constructing the SKB. The linear regression model between DI (%) and PRI at anthesis stage could be illustrated as follows:

**2.3.1.9 Verification** 

**2.3.2 Results** 

corresponding satellite images.

1. Verification of SKB using simulated data

2. Verification of SKB using field surveyed data

#### **2.3.1.8 Spectral matching algorithms**

The basic idea of spectral matching is to identify a set of pixels in the SKB that are the closest to the undefined pixel in terms of spectral characteristics. Before matching, each pixel should be standardized to eliminate systematic variation caused by aerosol conditions or other factors as follows:

$$\mathcal{R}\_{\text{nor}} = \frac{R - R\_{\text{min}}}{R\_{\text{max}} - R\_{\text{min}}}$$

where Rnor is the standardized reflectance of a certain band, R is the original reflectance, and Rmin and Rmax are the minimum and maximum band reflectance values, respectively, of the corresponding pixel.

Mahalanobis distances (Mah) and Spectral angle (SA) were selected as the distance measurement criterion. Both types of distance measurements had been proved to be with high efficiency in reflecting the spectral discrepancy (South et al., 2004; Goovaerts et al., 2005; Becker et al., 2007). The Mah distance can be written as:

$$D\_{\mathcal{M}}(\mathbf{x}) = \sqrt{(\mathbf{x} - \mathbf{x}\_{\mathcal{R}}) \sum^{-1} (\mathbf{x} - \mathbf{x}\_{\mathcal{R}})^T} \quad \mathbf{x} = (\mathbf{x}\_{\mathcal{I}}, \mathbf{x}\_{\mathcal{Z}}, \mathbf{x}\_{\mathcal{R}}, \mathbf{x}\_{\mathcal{I}}), \ \mathbf{x}\_{\mathcal{R}} = \begin{pmatrix} \mathbf{x}\_{\mathcal{R}1} \ \mathbf{x}\_{\mathcal{R}2} \ \mathbf{x}\_{\mathcal{R}3} \ \mathbf{x}\_{\mathcal{R}4} \end{pmatrix}$$

where x1-4 are the reflectance of the pixel under test in band1 to band4, respectively; xR1-4 are the simulated reflectance of a specific pixel in SKB. ∑ is the covariance matrix between x and xR.SA can be calculated by the following formula:

$$\theta = \arccos \frac{\sum\_{i=1}^{4} \mathbf{x}\_i \mathbf{x}\_{Ri}}{\sqrt{\sum\_{i=1}^{4} \mathbf{x}\_i^2} \sqrt{\sum\_{i=1}^{4} \mathbf{x}\_{Ri}^2}} \quad \theta \in \left[0, \frac{\pi}{2}\right]^2$$

To determine the DI (%) or class of disease severity of an undened pixel, we have to calculate the Mah and spectral angle from this pixel to each pixel or class in the SKB. A longer distance or larger angle indicates that the pixel deviated from the undefined pixel, whereas a shorter distance or smaller angle indicates that it is similar to the undefined pixel. By selecting the most similar pixel, the severity class of an undefined pixel can be determined. To determine the DI (%) of a certain pixel, the weighted average method was used. According to the distance criteria above, the five most similar pixels were selected from the SKB. For each band of these pixels (here we used the hyperspectral bands extracted from the PHI image), the reflectance was processed according to the following equation:

$$R\_E = \frac{\sum\_{i=1}^{k} R\_i \times \frac{1}{d\_i}}{\sum\_{i=1}^{k} \frac{1}{d\_i}}$$

where *RE* is the estimated reflectance of a certain pixel through k-NN estimation; *Ri* is the reflectance of the *i*th nearest pixel according to the ranking order of the distance; *di* is the distance between the pixel under test to the *i*th nearest pixel.

#### **2.3.1.9 Verification**

52 Remote Sensing – Applications

The basic idea of spectral matching is to identify a set of pixels in the SKB that are the closest to the undefined pixel in terms of spectral characteristics. Before matching, each pixel should be standardized to eliminate systematic variation caused by aerosol conditions or

> *R R R R*

where Rnor is the standardized reflectance of a certain band, R is the original reflectance, and Rmin and Rmax are the minimum and maximum band reflectance values, respectively, of the

Mahalanobis distances (Mah) and Spectral angle (SA) were selected as the distance measurement criterion. Both types of distance measurements had been proved to be with high efficiency in reflecting the spectral discrepancy (South et al., 2004; Goovaerts et al.,

where x1-4 are the reflectance of the pixel under test in band1 to band4, respectively; xR1-4 are the simulated reflectance of a specific pixel in SKB. ∑ is the covariance matrix between x and

*i Ri*

*x x*

*i Ri i i*

0, 2 

*x x*

To determine the DI (%) or class of disease severity of an undened pixel, we have to calculate the Mah and spectral angle from this pixel to each pixel or class in the SKB. A longer distance or larger angle indicates that the pixel deviated from the undefined pixel, whereas a shorter distance or smaller angle indicates that it is similar to the undefined pixel. By selecting the most similar pixel, the severity class of an undefined pixel can be determined. To determine the DI (%) of a certain pixel, the weighted average method was used. According to the distance criteria above, the five most similar pixels were selected from the SKB. For each band of these pixels (here we used the hyperspectral bands extracted from the PHI image), the reflectance was processed according to the following equation:

1

 

*R <sup>d</sup> <sup>R</sup>*

*k i i i E k*

1

where *RE* is the estimated reflectance of a certain pixel through k-NN estimation; *Ri* is the reflectance of the *i*th nearest pixel according to the ranking order of the distance; *di* is the

*i i*

1

*d*

1

4

1 4 4 2 2 1 1

*i*

arccos

x=(x1,x2,x3,x4), xR=(xR1, xR2, xR3, xR4)

R*nor*

2005; Becker et al., 2007). The Mah distance can be written as:

<sup>1</sup> () ( ) ( )*<sup>T</sup> D x xx xx M RR*

distance between the pixel under test to the *i*th nearest pixel.

xR.SA can be calculated by the following formula:

min max min

**2.3.1.8 Spectral matching algorithms** 

other factors as follows:

corresponding pixel.

To verify the performance of SKB in identifying and monitoring the severity of yellow rust diseases, two datasets were used: the simulated data and the field-measured data with corresponding satellite images.

1. Verification of SKB using simulated data

The simulated data comprised 50 randomly selected pixels in the same experimental field, but outside the three regions selected for constructing the SKB. The hyperspectral information of each pixel was used to create the reference DI (%) and severity class with the empirical model and the corresponding threshold for each severity class. To test the performance of SKB in terms of DI (%) value, we estimated the DI value with both distance criteria described above. The samples were split into two: the pixels with a reference DI between 1 and 100%, i.e. the 'diseased' pixels, and those with a reference DI<1%, i.e. 'healthy' pixels. For the diseased pixels, the estimated DIs were compared with the reference DI by Pearson correlation analysis and the normalized root mean square error (NRMSE). For the healthy pixels, we used 'yes or no' to determine whether the estimated value indicated infection or not, which also provided an accuracy ratio. The estimation of severity class was verified by overall accuracy and the kappa coefficient.

2. Verification of SKB using field surveyed data The field surveyed data sets included the ground investigation of disease severity and the corresponding HJ-CCD images. Between June 1–3, 2009, when the winter wheat was at the anthesis stage, we conducted a survey in the southeast of GanSu Province. The climate of the area surveyed is characterized by high humidity and rainfall, and yellow rust disease occurs almost every year. This area has similar environmental conditions and cultivation customs to those where we constructing the SKB in Beijing, and this makes it an appropriate place for model verification. With the aid of the local Department of Plant Protection, 26 plots were randomly selected and surveyed in the area (Fig. 9). To relate the surveyed value to the pixel value of the HJ-CCD image, we defined the plot as a uniformly planted winter wheat region with an area no less than 30 m in radius. The geographical coordinates of each plot were measured by GPS at the centre of the plot. Disease severity was measured as described above. We repeated the measurement in ve evenlydistributed sections in each plot, and 20 individual plants were included in each measurement. The HJ-CCD images (ID: 122516, 122518) acquired on June 2, 2009 completely covered the surveyed area. The raw data from the HJ-CCD imagery was calibrated based on the corresponding coefficients provided by CRESDA. The calibrated data were atmospherically corrected with the algorithm provided by Liang et al. (2001), which estimated the spatial distribution of atmospheric aerosols and retrieved surface reflectance under general atmospheric and surface conditions. The images were also geometrically corrected against historical reference images with the same geographical coordinates. The images were rectified with a root mean square error of less than 0.5 pixels. The spectrum of the each plot was extracted from the image according to the GPS records. The estimated accuracy in this step followed the same process as the simulated data.

#### **2.3.2 Results**

There were 7918 pixels included in the process of constructing the SKB. The linear regression model between DI (%) and PRI at anthesis stage could be illustrated as follows:

Crop Disease and Pest Monitoring by Remote Sensing 55

Table 6 gives the reference class of disease severity and the estimated class in the form of an error matrix. The overall accuracy with Mah distance and the SA criterion were 0.80 and 0.76, respectively, whereas the kappa coefficients were 0.71 and 0.65, respectively. However, we noticed that all the misclassified pixels were assigned to no more than one class adjacent to the reference class. Therefore, for simulated data, the classification accuracy was

Apart from the verification against simulated data, more importantly, the field surveyed data can be also used to assess the performance of the SKB. The field investigation showed that eight out of 26 plots were infected with DI ranged from 4 to 90%, whereas the other 18 plots were not affected by yellow rust. The estimation by DI (%) successfully identified the eight infected plots when the Mah distance criterion was used, whereas the SA criterion

Fig. 10. MNF eigenvalues variation trend

Fig. 11. Estimated DI(%) using simulated data

**2.3.2.2 Performance of SKB for field surveyed data** 

satisfactory in determining the severity class of yellow rust by SKB.

Fig. 9. The field surveyed area in Gansu Province. The base image is the HJ-CCD image acquired on June 2, 2009

$$DI(\%) = -538.98 \times PRI + 2.0983 \text{ (R $^2$ =0.88)}$$

The pairs of DI (%) and PRI were plotted in Fig.4, which showed a significant correlation (R2 = 0.88). Based on the model, there were 85 pixels with a DI of 100% and 3991 pixels with a DI between 1%and 100%, indicating 51.5%pixels infected to a varied degree of severity, whereas the other 48.5% pixels (DI = 0%) were healthy plants. In the experimental field, the variation in the degree of severity of yellow rust from totally healthy plants to very infected plants provided the essential diversity or heterogeneity of infection, which then enabled establishment of the SKB. The MNF transformation resulted in 9 leading eigenvectors with eigenvalues greater than 4.0 (Fig. 10), and these were used for further analysis.

#### **2.3.2.1 Performance of SKB for simulated data**

In the simulated dataset, there were six healthy pixels and 44 diseases affected ones. When estimating DI (%), one pixel with no infection was estimated as infected by the Mah distance criterion, whereas with the SA criterion two were mislabeled. Fig.11 shows the scatter of the disease affected pixels plotted in relation to reference DI and estimated DI; the average reference DI is 36%. The reference DIs and estimated DIs were strongly and linearly correlated for both the Mah distance (R2 = 0.90) and SA (R2 = 0.84) criteria. Further, the NRMSE of Mah distance and SA were 0.20 and 0.24, respectively, indicating that the SKB can estimate DIs accurately from the simulated multi-band reflectance.

Fig. 9. The field surveyed area in Gansu Province. The base image is the HJ-CCD image

*DI*(%) 538.98 2.0983 *PRI* (R2=0.88)

The pairs of DI (%) and PRI were plotted in Fig.4, which showed a significant correlation (R2 = 0.88). Based on the model, there were 85 pixels with a DI of 100% and 3991 pixels with a DI between 1%and 100%, indicating 51.5%pixels infected to a varied degree of severity, whereas the other 48.5% pixels (DI = 0%) were healthy plants. In the experimental field, the variation in the degree of severity of yellow rust from totally healthy plants to very infected plants provided the essential diversity or heterogeneity of infection, which then enabled establishment of the SKB. The MNF transformation resulted in 9 leading eigenvectors with

In the simulated dataset, there were six healthy pixels and 44 diseases affected ones. When estimating DI (%), one pixel with no infection was estimated as infected by the Mah distance criterion, whereas with the SA criterion two were mislabeled. Fig.11 shows the scatter of the disease affected pixels plotted in relation to reference DI and estimated DI; the average reference DI is 36%. The reference DIs and estimated DIs were strongly and linearly correlated for both the Mah distance (R2 = 0.90) and SA (R2 = 0.84) criteria. Further, the NRMSE of Mah distance and SA were 0.20 and 0.24, respectively, indicating that the SKB

eigenvalues greater than 4.0 (Fig. 10), and these were used for further analysis.

can estimate DIs accurately from the simulated multi-band reflectance.

**2.3.2.1 Performance of SKB for simulated data** 

acquired on June 2, 2009

Fig. 10. MNF eigenvalues variation trend

Fig. 11. Estimated DI(%) using simulated data

Table 6 gives the reference class of disease severity and the estimated class in the form of an error matrix. The overall accuracy with Mah distance and the SA criterion were 0.80 and 0.76, respectively, whereas the kappa coefficients were 0.71 and 0.65, respectively. However, we noticed that all the misclassified pixels were assigned to no more than one class adjacent to the reference class. Therefore, for simulated data, the classification accuracy was satisfactory in determining the severity class of yellow rust by SKB.

#### **2.3.2.2 Performance of SKB for field surveyed data**

Apart from the verification against simulated data, more importantly, the field surveyed data can be also used to assess the performance of the SKB. The field investigation showed that eight out of 26 plots were infected with DI ranged from 4 to 90%, whereas the other 18 plots were not affected by yellow rust. The estimation by DI (%) successfully identified the eight infected plots when the Mah distance criterion was used, whereas the SA criterion

Crop Disease and Pest Monitoring by Remote Sensing 57

determining the disease severity. This might be because it is more rough estimation than DI (%). It is understandable that for the same sample, the less precise the criterion, the greater accuracy it would achieve. Moreover, the 5-class disease severity quantification is enough to guide field applications. We suggest that DI (%) should be used for detecting the disease severity of yellow rust by SKB. For the distance criteria used in the process of matching with SKB, the Mah distance criterion might be more appropriate because it performed better than SA in all the analyses conducted in this study (Figs. 11, 12, Tables 6, 7). Some previous studies have already emphasized the potential of hyperspectral imagery (Bravo et al. 2003; Moshou et al. 2004; Huang et al. 2007) and the high-resolution of multispectral imagery (Franke and Menz 2007) for detecting yellow rust disease. The development of SKB in the present study can be viewed as a scaling up method, which has extended the capability of detecting yellow rust disease from hyper- spectral imagery to the moderate resolution of multispectral imagery. However, it should be noted that the task of monitoring the occurrence and degrees of infection of crop diseases is far more complex than the cases described in this study. The spectral characteristics of yellow rust infection might appear similar to other sources of stress. In addition, the impact of phenology, cultivation methods, fragmentation of farmlands and other environmental conditions would also increase the difficulty and uncertainty of the estimation process. Therefore, the SKB developed in this study should correspond to the situation at the anthesis stage exclusively, and is only suitable for those regions with similar environmental characteristics and cultivation methods. For other regions with significantly different environmental characteristics, this purposed SKB may not work well. The possible solution to these problems may include incorporating suitable priors, which would require integration strategies and understanding of the mechanisms underlying some fundamental processes.

Further research is required to address the problems mentioned above.

Fig. 12. Estimated DI(%) using field measurements

The low spatial resolution and few spectral bands have limited the application of moderate resolution satellite images for monitoring yellow rust disease. The spectral knowledge base developed enabled disease incidence and severity to be detected by moderate resolution satellite images. The SKB supported two ways of estimating disease severity: the disease

**2.3.3 Conclusion** 


Table 6. Error matrix for simulated data

resulted in one misestimated plot. Figure 7 shows the scatter of the eight data plotted in relation to reference DI and estimated DI for both distance criteria. There was a significant linear trend in graphs based on both the Mah distance and SA criteria. The R2 of Mah distance and SA were 0.80 and 0.67, respectively, whereas the NRMSE were as high as 0.46 and 0.55. In real circumstances, approximately 50% error in the estimated disease index is unsatisfactory. On the other hand, however, most of the uninfected plots were correctly identified according to DI (%) estimates (i.e. a DI<1%). For both the Mah distance and SA criteria, 15 out of 18 noninfected plots had been identified correctly, resulting in an accuracy of 77.8%. The results for estimating disease severity by severity class were even more encouraging. The overall accuracy for the Mah distance and SA criteria were 0.77 and 0.73, respectively, whereas the kappa coefficients are 0.58 and 0.49, respectively. Table 3 gives the error matrix for both criteria. The misclassified pixels were also assigned exclusively to the adjacent class.

In general, the above results demonstrate that the proposed SKB scheme has great potential for detecting the incidence and severity of yellow rust through multispectral images. As shown from several previous studies, the image processing method of MNF transformation was efficient in extracting the principle information from the images related to wheat disease infection (Zhang et al. 2003; Franke and Menz 2007). For the present study, we found that coupling MNF transformation with the PPI function was an appropriate way of extracting the principle information on yellow rust disease. To estimate disease severity by DI (%), the proposed SKB has achieved a satisfactory accuracy for simulated data. However, the estimated accuracy for field surveyed data was unsatisfactory, implying that the method tends to underestimate or overestimate the disease severity in practice. Nevertheless, to estimate disease severity through disease severity class has achieved a satisfactory accuracy for both simulated data and field surveyed data. Therefore, the disease severity class seems to be more robust in

Mid range

None 6 0 0 0 0 6 Low range 0 5 2 0 0 7 Mid range 0 1 20 2 0 23 Serious 0 0 1 10 1 12

serious 0 0 0 1 1 2 Total 6 6 23 13 2 50

None 5 1 0 0 0 6 Low range 1 4 1 0 0 6 Mid range 0 1 20 2 0 23 Serious 0 0 2 9 1 12

serious 0 0 0 2 1 3 Total 6 6 23 13 2 50

resulted in one misestimated plot. Figure 7 shows the scatter of the eight data plotted in relation to reference DI and estimated DI for both distance criteria. There was a significant linear trend in graphs based on both the Mah distance and SA criteria. The R2 of Mah distance and SA were 0.80 and 0.67, respectively, whereas the NRMSE were as high as 0.46 and 0.55. In real circumstances, approximately 50% error in the estimated disease index is unsatisfactory. On the other hand, however, most of the uninfected plots were correctly identified according to DI (%) estimates (i.e. a DI<1%). For both the Mah distance and SA criteria, 15 out of 18 noninfected plots had been identified correctly, resulting in an accuracy of 77.8%. The results for estimating disease severity by severity class were even more encouraging. The overall accuracy for the Mah distance and SA criteria were 0.77 and 0.73, respectively, whereas the kappa coefficients are 0.58 and 0.49, respectively. Table 3 gives the error matrix for both

criteria. The misclassified pixels were also assigned exclusively to the adjacent class.

In general, the above results demonstrate that the proposed SKB scheme has great potential for detecting the incidence and severity of yellow rust through multispectral images. As shown from several previous studies, the image processing method of MNF transformation was efficient in extracting the principle information from the images related to wheat disease infection (Zhang et al. 2003; Franke and Menz 2007). For the present study, we found that coupling MNF transformation with the PPI function was an appropriate way of extracting the principle information on yellow rust disease. To estimate disease severity by DI (%), the proposed SKB has achieved a satisfactory accuracy for simulated data. However, the estimated accuracy for field surveyed data was unsatisfactory, implying that the method tends to underestimate or overestimate the disease severity in practice. Nevertheless, to estimate disease severity through disease severity class has achieved a satisfactory accuracy for both simulated data and field surveyed data. Therefore, the disease severity class seems to be more robust in

Serious Very

serious Total

Reference

None Low range

Estimation (Mah)

Estimation (SA)

Very

Very

Table 6. Error matrix for simulated data

determining the disease severity. This might be because it is more rough estimation than DI (%). It is understandable that for the same sample, the less precise the criterion, the greater accuracy it would achieve. Moreover, the 5-class disease severity quantification is enough to guide field applications. We suggest that DI (%) should be used for detecting the disease severity of yellow rust by SKB. For the distance criteria used in the process of matching with SKB, the Mah distance criterion might be more appropriate because it performed better than SA in all the analyses conducted in this study (Figs. 11, 12, Tables 6, 7). Some previous studies have already emphasized the potential of hyperspectral imagery (Bravo et al. 2003; Moshou et al. 2004; Huang et al. 2007) and the high-resolution of multispectral imagery (Franke and Menz 2007) for detecting yellow rust disease. The development of SKB in the present study can be viewed as a scaling up method, which has extended the capability of detecting yellow rust disease from hyper- spectral imagery to the moderate resolution of multispectral imagery. However, it should be noted that the task of monitoring the occurrence and degrees of infection of crop diseases is far more complex than the cases described in this study. The spectral characteristics of yellow rust infection might appear similar to other sources of stress. In addition, the impact of phenology, cultivation methods, fragmentation of farmlands and other environmental conditions would also increase the difficulty and uncertainty of the estimation process. Therefore, the SKB developed in this study should correspond to the situation at the anthesis stage exclusively, and is only suitable for those regions with similar environmental characteristics and cultivation methods. For other regions with significantly different environmental characteristics, this purposed SKB may not work well. The possible solution to these problems may include incorporating suitable priors, which would require integration strategies and understanding of the mechanisms underlying some fundamental processes. Further research is required to address the problems mentioned above.

Fig. 12. Estimated DI(%) using field measurements

#### **2.3.3 Conclusion**

The low spatial resolution and few spectral bands have limited the application of moderate resolution satellite images for monitoring yellow rust disease. The spectral knowledge base developed enabled disease incidence and severity to be detected by moderate resolution satellite images. The SKB supported two ways of estimating disease severity: the disease

Crop Disease and Pest Monitoring by Remote Sensing 59

district, Tianshui district, Dingxi district and Pingliang district in GanSu province and Qingyang district in ShanXi province as well as Linxia district in Ningxia Hui Autonomous Region (Fig.1), where the climates are semiarid and subhumid. Survey areas are located between latitude 32º40'N to 35º39'N and longitude 103º10'E to 107º40'E, and the mean altitude is over 2000 meter. The climate condition of surveyed area is characterized by high humidity and rainfall, and yellow rust disease almost occurs every year. It is reported that Longnan district is an important overwintering and oversummering area of yellow rust

With the aid of the local Department of Plant Protection, 151 plots, including 68 plots from April to June in 2008, and 83 plots from April to June in 2009, were randomly selected and surveyed in the areas. The geographical coordinates of each plot were measured by GPS navigator at the middlemost of the plot. In addition, the disease severity was inspected.

MODIS Land Surface Temperature and Emissivity (LST/E) products (named starting with MOD11) provide per-pixel temperature and emissivity values. Temperatures are extracted in Kelvin with a view-angle dependent algorithm applied to direct observations. This method yields the error less than 1 K for materials with known emissivity. The view angle

24 MOD11A2 images(MODIS/Terra land surface temperature/emissivity 8-day L3 global 1km SIN grid v005)were acquired for free from Web (http://edc.usgs.gov/#/Find\_Data) from April to July in 2008 and 2009, which covered completely the survey area, and 4 scenes images were acquired in every month. The raw data of MOD11A2 imagery were processed and transformed by MRT tool, and LST products were extracted from MODII A2 images. Then the survey area was cut by ENVI from LST images. Followed by that step, 4 scenes 8 day LST images of every month were all averaged, and 6 average LST images, including April, May, June in 2008 and 2009, were obtained. Finally, LST of 151 investigation points

The spatial resolution of MODIS temperature products is 1 km, while the DI of every investigation point only stands for the incidence of 30 m in semi diameter plots. Therefore, the scale of MODIS temperature products seemed not satisfied the investigation points for proper relationship between them. However, spatial variability of LST is slim, and the law still exists. A series of results could be found by establishing a two-dimensional spatial coordinate based on DI and LST, in which all investigation points were displayed (Fig 13). Firstly, the DI ranged from 0% to 100%, and most of infected points ranged from 0% to 60%. The LST values were between 292K and 310K with most of infected points distributed in the range from 298K to 306K. In addition, the points in the region of less than 298K were not infected by yellow rust basically; DI were less than 1% expect for one point (296.29K, 16%),

**2.4.1.2 MODIS land surface temperature (LST) products (MOD11)** 

information is included in each LST/E product.

were respectively extracted from 6 average LST images.

**2.4.2.1 Determining LST threshold of infected points** 

**MOD11 acquisition and processing** 

fungal (Zeng, 2003).

**Product description** 

**2.4.2 Result** 


Table 7. Error matrix for ground measured data

index and disease severity class. Both methods of estimation achieved a satisfactory level of accuracy for simulated data. For field surveyed data, estimation by DI (%) resulted in an unsatisfactory level of accuracy, whereas it was satisfactory for severity class. The Mah criterion performed better than spectral angle in all analyses. Therefore, the former should be considered as the more appropriate distance criterion.

Generally, the purposed SKB has a great potential in extending the capability of detecting yellow rust to multispectral remote sensing data, especially when the region of interest has similar environmental conditions to where the SKB was developed. The uncertainties caused by environmental differences should be further investigated in future studies.

#### **2.4 Detecting yellow rust of winter wheat using land surface temperature (LST)**

The air temperature and humidity are the most direct and important indicators of occurrence of yellow rust fungal. Generally, weather stations can provide the dynamic pattern of meteorological data for site sampled, yet not able to include the information of spatial heterogeneity. Fortunately, remote sensing technology has great potential for providing spatially continuous observations of some variables over large areas (Luo et al., 2010). The aim of the study was to study preliminarily on the relationship between the occurrence of wheat yellow rust and land surface temperature (LST) derived from moderate-resolution imaging spectroradiometer (MODIS) in order to predict and monitor incidence of the yellow rust on large scale.

#### **2.4.1 Materials and methods**

#### **2.4.1.1 Survey area and field investigations acquisition**

Field experiments of winter wheat were conducted during the growing seasons (form April to June) of winter wheat in 2008 and 2009. The investigation locations included Longnan district, Tianshui district, Dingxi district and Pingliang district in GanSu province and Qingyang district in ShanXi province as well as Linxia district in Ningxia Hui Autonomous Region (Fig.1), where the climates are semiarid and subhumid. Survey areas are located between latitude 32º40'N to 35º39'N and longitude 103º10'E to 107º40'E, and the mean altitude is over 2000 meter. The climate condition of surveyed area is characterized by high humidity and rainfall, and yellow rust disease almost occurs every year. It is reported that Longnan district is an important overwintering and oversummering area of yellow rust fungal (Zeng, 2003).

With the aid of the local Department of Plant Protection, 151 plots, including 68 plots from April to June in 2008, and 83 plots from April to June in 2009, were randomly selected and surveyed in the areas. The geographical coordinates of each plot were measured by GPS navigator at the middlemost of the plot. In addition, the disease severity was inspected.

#### **2.4.1.2 MODIS land surface temperature (LST) products (MOD11)**

#### **Product description**

58 Remote Sensing – Applications

None Low range Mid range Serious Very serious Total

None 16 0 0 0 0 16 Low range 2 2 1 0 0 5 Mid range 0 1 3 0 0 4 Serious 0 0 0 0 1 1 Very serious 0 0 0 0 0 0 Total 18 3 4 0 1 26

None 15 0 0 0 0 15 Low range 3 2 1 0 0 6 Mid range 0 1 3 0 0 4 Serious 0 0 0 0 1 1 Very serious 0 0 0 0 0 0 Total 18 3 4 0 1 26

index and disease severity class. Both methods of estimation achieved a satisfactory level of accuracy for simulated data. For field surveyed data, estimation by DI (%) resulted in an unsatisfactory level of accuracy, whereas it was satisfactory for severity class. The Mah criterion performed better than spectral angle in all analyses. Therefore, the former should

Generally, the purposed SKB has a great potential in extending the capability of detecting yellow rust to multispectral remote sensing data, especially when the region of interest has similar environmental conditions to where the SKB was developed. The uncertainties caused by environmental differences should be further investigated in future studies.

The air temperature and humidity are the most direct and important indicators of occurrence of yellow rust fungal. Generally, weather stations can provide the dynamic pattern of meteorological data for site sampled, yet not able to include the information of spatial heterogeneity. Fortunately, remote sensing technology has great potential for providing spatially continuous observations of some variables over large areas (Luo et al., 2010). The aim of the study was to study preliminarily on the relationship between the occurrence of wheat yellow rust and land surface temperature (LST) derived from moderate-resolution imaging spectroradiometer (MODIS) in order to predict and monitor

Field experiments of winter wheat were conducted during the growing seasons (form April to June) of winter wheat in 2008 and 2009. The investigation locations included Longnan

**2.4 Detecting yellow rust of winter wheat using land surface temperature (LST)** 

Reference

Estimation (Mah)

Estimation (SA)

Table 7. Error matrix for ground measured data

incidence of the yellow rust on large scale.

**2.4.1.1 Survey area and field investigations acquisition** 

**2.4.1 Materials and methods** 

be considered as the more appropriate distance criterion.

MODIS Land Surface Temperature and Emissivity (LST/E) products (named starting with MOD11) provide per-pixel temperature and emissivity values. Temperatures are extracted in Kelvin with a view-angle dependent algorithm applied to direct observations. This method yields the error less than 1 K for materials with known emissivity. The view angle information is included in each LST/E product.

#### **MOD11 acquisition and processing**

24 MOD11A2 images(MODIS/Terra land surface temperature/emissivity 8-day L3 global 1km SIN grid v005)were acquired for free from Web (http://edc.usgs.gov/#/Find\_Data) from April to July in 2008 and 2009, which covered completely the survey area, and 4 scenes images were acquired in every month. The raw data of MOD11A2 imagery were processed and transformed by MRT tool, and LST products were extracted from MODII A2 images. Then the survey area was cut by ENVI from LST images. Followed by that step, 4 scenes 8 day LST images of every month were all averaged, and 6 average LST images, including April, May, June in 2008 and 2009, were obtained. Finally, LST of 151 investigation points were respectively extracted from 6 average LST images.

#### **2.4.2 Result**

#### **2.4.2.1 Determining LST threshold of infected points**

The spatial resolution of MODIS temperature products is 1 km, while the DI of every investigation point only stands for the incidence of 30 m in semi diameter plots. Therefore, the scale of MODIS temperature products seemed not satisfied the investigation points for proper relationship between them. However, spatial variability of LST is slim, and the law still exists. A series of results could be found by establishing a two-dimensional spatial coordinate based on DI and LST, in which all investigation points were displayed (Fig 13). Firstly, the DI ranged from 0% to 100%, and most of infected points ranged from 0% to 60%. The LST values were between 292K and 310K with most of infected points distributed in the range from 298K to 306K. In addition, the points in the region of less than 298K were not infected by yellow rust basically; DI were less than 1% expect for one point (296.29K, 16%),

Crop Disease and Pest Monitoring by Remote Sensing 61

Furthermore, there was an increasing trend of incidences with the rising of LST in the region from 296K to 302K. The incidence of yellow rust reached up to 100% when the LST was

According to Table 8 and Fig. 14, the survey areas could be divided into yellow rust unsuitable area (NSA), of which LST ranged from 298K to 306K, and yellow rust suitable area (SA), of which the LST was less than 298K and more than 306K. Moreover, the SA was divided into 3 levels according to the infected of yellow rust incidence and LST, and the LST thresholds for each level were: 298K ≤ LST ≤ 299K the low suitable area (LSA), on which the yellow rust occurs with very low possibility (incidence < 60%), 299K ≤ LST ≤ 301K the medium suitable area (MSA), which had moderate possibility for the occurrence of yellow rust (60% <incidence < 100%), and 302K ≤ LST ≤ 306K high suitable area (HSA), of which the

Total 26 points (from May 2008) were applied for the verification the method of estimating the incidence of yellow rust. It should be noted that those points were not used for the defining of the LST thresholds. (Fig. 15). These 26 points were constituted by 18 infected points and 8 noninfected points. Results showed the infected points were all in different suitable areas of wheat yellow rust, while the non-infected points were all in the unsuitable area. Thus the infected situation of yellow rust of these 26 points was consistent with forecast results. Geographically, it seemed that the yellow rust was prone to be prevalent in the northeast of Pingliang, southwest of Qingyang, northeast of Dingxi, the center part of Tianshui, and the west of Longnan, because they all were located in MSA and HAS. This result was consistent with the previous study (Xiao, et al, 2007). To prevent yellow rust from prevalence, more efforts should

Plant disease is governed by a number of factors, and the habitat factors play a major role in the development and propagation of fungal pathogens (Sutton et al., 1984; Hélène et al.,

graeter than 302K (Fig. 14).

**2.4.2.4 Verification** 

**2.4.3 Conclusions** 

Fig. 14. The incidence of yellow rust in different LST range

**2.4.2.3 Dividing yellow rust suitable occurrence region based on LST** 

environment was highly favorable to yellow rust (incidence=100%).

be placed on the farmlands located in the MSA, HAS and LSA.

which was thought as abnormal point. In addition, the LST values of all investigation points were less than 306K expect for one point (310.09K, 24%), which was abnormal because its LST was far away from LST values of others.

Fig. 13. The distribution of the investigation points

Therefore, without considering other factors, It is concluded that yellow rust can occur when LST is in the region from 298K to 306K.

#### **2.4.2.2 Yellow rust incidence analysis based on LST**

According to the results illustrated above, the advanced analysis was performed for incidence and possible area of yellow rust. The points in different LST range were done statistical analysis with all points' numbers and the infected points' number, and finally, the incidences were obtained by the number of the infected points dividing the number of all points in the different LST range (Table.8). The result showed that all investigation points in the region of less than 298K were not infected by yellow rust, except for the abnormal point (296.29K, 16%). On the other hand, in the LST region of more than 306K, there was only one point, which was viewed as abnormal point (310.085K, 24%). Thereby, it is quite possible that yellow rust fungus can not survive in the region of more than 306K. The conclusion was consistent with the above result (Fig. 13).


Table 8. Statistic analysis in different LST range

which was thought as abnormal point. In addition, the LST values of all investigation points were less than 306K expect for one point (310.09K, 24%), which was abnormal because its

Therefore, without considering other factors, It is concluded that yellow rust can occur

According to the results illustrated above, the advanced analysis was performed for incidence and possible area of yellow rust. The points in different LST range were done statistical analysis with all points' numbers and the infected points' number, and finally, the incidences were obtained by the number of the infected points dividing the number of all points in the different LST range (Table.8). The result showed that all investigation points in the region of less than 298K were not infected by yellow rust, except for the abnormal point (296.29K, 16%). On the other hand, in the LST region of more than 306K, there was only one point, which was viewed as abnormal point (310.085K, 24%). Thereby, it is quite possible that yellow rust fungus can not survive in the region of more than 306K. The conclusion was

LST was far away from LST values of others.

Fig. 13. The distribution of the investigation points

**2.4.2.2 Yellow rust incidence analysis based on LST** 

when LST is in the region from 298K to 306K.

consistent with the above result (Fig. 13).

LST≥2 97

Table 8. Statistic analysis in different LST range

LST≥ 298

LST≥ 299

LST≥ 300

number 126 112 99 79 61 34 25 16 12 8 1

Incidence (%) 38.89 42.86 47.47 53.16 63.93 79.41 100 100 100 100 100

LST≥ 301

49 48 47 42 39 27 25 16 12 8 1

LST≥ 302

LST≥ 303

LST≥ 304

LST≥ 305

LST≥ 306

96

LST (K) LST≥<sup>2</sup>

Total

Number of infected points

Furthermore, there was an increasing trend of incidences with the rising of LST in the region from 296K to 302K. The incidence of yellow rust reached up to 100% when the LST was graeter than 302K (Fig. 14).

Fig. 14. The incidence of yellow rust in different LST range

#### **2.4.2.3 Dividing yellow rust suitable occurrence region based on LST**

According to Table 8 and Fig. 14, the survey areas could be divided into yellow rust unsuitable area (NSA), of which LST ranged from 298K to 306K, and yellow rust suitable area (SA), of which the LST was less than 298K and more than 306K. Moreover, the SA was divided into 3 levels according to the infected of yellow rust incidence and LST, and the LST thresholds for each level were: 298K ≤ LST ≤ 299K the low suitable area (LSA), on which the yellow rust occurs with very low possibility (incidence < 60%), 299K ≤ LST ≤ 301K the medium suitable area (MSA), which had moderate possibility for the occurrence of yellow rust (60% <incidence < 100%), and 302K ≤ LST ≤ 306K high suitable area (HSA), of which the environment was highly favorable to yellow rust (incidence=100%).

#### **2.4.2.4 Verification**

Total 26 points (from May 2008) were applied for the verification the method of estimating the incidence of yellow rust. It should be noted that those points were not used for the defining of the LST thresholds. (Fig. 15). These 26 points were constituted by 18 infected points and 8 noninfected points. Results showed the infected points were all in different suitable areas of wheat yellow rust, while the non-infected points were all in the unsuitable area. Thus the infected situation of yellow rust of these 26 points was consistent with forecast results. Geographically, it seemed that the yellow rust was prone to be prevalent in the northeast of Pingliang, southwest of Qingyang, northeast of Dingxi, the center part of Tianshui, and the west of Longnan, because they all were located in MSA and HAS. This result was consistent with the previous study (Xiao, et al, 2007). To prevent yellow rust from prevalence, more efforts should be placed on the farmlands located in the MSA, HAS and LSA.

#### **2.4.3 Conclusions**

Plant disease is governed by a number of factors, and the habitat factors play a major role in the development and propagation of fungal pathogens (Sutton et al., 1984; Hélène et al.,

Crop Disease and Pest Monitoring by Remote Sensing 63

occurrence and prevalence of aphid, large amounts of insecticides are used, causing environment pollution. Therefore, large-scale, real-time prediction and monitoring of wheat aphid incidence and damage degree using remote sensing technology are extremely

The study aimed to identify spectral characteristics of wheat leaf and canopy infected by aphid and find the sensitive bands to aphid at canopy level in filling stage of wheat, and to establish an aphid damage hyperspectral index (ADHI) based on those sensitive bands for

The field experiment plot was located at Xiaotangshan Precision Agriculture Experiment Base, Changping distract, Beijing (40º10.6'N, 116º26.3'E). The experimental field was about 250 m in length and 80 m in width. The winter wheat was planted in the study area from Oct 3, 2009, and harvested from June 25, 2010. Field inventory was conducted on June 7, 2010 when wheat was in the filling stage. Twenty five ground investigations including different aphid damage levels were selected. Aphid damage level was surveyed according

Representative reflectance measured from wheat aphid-infested and uninfested wheat leaves are shown in Fig. 16. It was evident that the spectral response of the wheat leaf was significantly affected by wheat aphid feeding (Fig. 16). The reflectance of wheat leaf infested by aphid was higher in the visible spectrum and short-wave infrared region and lower in near-infrared region than that of uninfested leaf. A significant increase in the reflectance from the wheat aphid-infested leaf in the visible region (400-700 nm) was observed, evidently due to reduction of photosynthetic pigment concentrations in particular

Compared with the canopy spectra of the healthy wheat, the canopy reflectance of aphidinfested wheat was gradually decreased in the range from 350 nm to 1750 nm, especially in the near infrared region (Fig. 17). Previous researches indicated that wheat had higher reflectance at visible wavelengths than the healthy vigorously growing wheat because the photoactive pigments (chlorophylls, anthocyanins, carotenoids) were destroyed. In this study, aphid occurred in the filling stage of wheat and the honeydew excreted by aphid absorbed dust or others from surrounding environment and contaminated (darkened) the leaf surface. As a

result, the absorption at light slight wavelengths became stronger instead of weaker.

detecting aphid damage levels in wheat canopy level in filling stage of wheat.

**3.1 Detecting winter wheat aphid using hyperspectral data** 

**3.1.2.1 Leaf spectral characteristics of wheat infested by aphid** 

chlorophylls caused by wheat aphid feeding (Richardson et al., 2004). **3.1.2.2 Canopy spectral characteristics of wheat infested by aphid** 

important.

**3.1.1 Materials and methods** 

to the investigation rule.

**3.1.2 Results** 

**3.1.1.2 Canopy spectral measurements** 

Please refer to 1.1.1.2 part above.

**3.1.1.1 Field experiments and field inventory**

Fig. 15. Forecast map of yellow rust and distribution of measured points in May, 2008 based on LST

2002; Cooke et al., 2006). The yellow rust is no exception. The weather station can only offer points data, and remote sensing, however, can be a promising means for acquiring spatially continuous observations over large area. It has not been reported, if any, that the LST derived from remote sensing data is used to forecast the development of yellow rust.

The study tried to present a method that could forecast the suitable areas of wheat yellow rust by MODIS temperature products in a large scale. And it was proved that LST derived from remote sensing data had potential for predicting the occurrence and development of wheat yellow rust in a large area. From our results, it is clear that preventive measures of yellow rust can been made over large scale area accordingly with different real-time prediction methods based on LST derived from remote sensing data.

#### **3. Detecting and discriminating winter wheat aphid by remote sensing**

Wheat aphid, Sitobion avenae *(Fabricius)*, is one of the most destructive pests in agricultural systems, especially in temperate climates of the northern and southern hemispheres. Wheat aphid appears annually in the wheat planting area of China, causing great economic damage to plant crops as a result of their direct feeding activities. In high enough densities, wheat aphids can remove plant nutrients, and potentially reduce the number of heads, the number of grains per head, and overall seed weight. The damage is especially high when wheat aphid occurs in the flowering and filling stage of wheat. It is reported that average densities over 20 aphids per plant can cause substantial losses of yield and quality of wheat (Basky & Fónagy, 2003). There are also indirect damages including excretion of honeydew from aphids and as a vector of viruses, most notably two strains of the Luteovirus Barley Yellow Dwarf Virus (BYDV-MAV and BYDV-PAV) (Susan et al, 1992). To prevent the occurrence and prevalence of aphid, large amounts of insecticides are used, causing environment pollution. Therefore, large-scale, real-time prediction and monitoring of wheat aphid incidence and damage degree using remote sensing technology are extremely important.

### **3.1 Detecting winter wheat aphid using hyperspectral data**

The study aimed to identify spectral characteristics of wheat leaf and canopy infected by aphid and find the sensitive bands to aphid at canopy level in filling stage of wheat, and to establish an aphid damage hyperspectral index (ADHI) based on those sensitive bands for detecting aphid damage levels in wheat canopy level in filling stage of wheat.

#### **3.1.1 Materials and methods**

#### **3.1.1.1 Field experiments and field inventory**

The field experiment plot was located at Xiaotangshan Precision Agriculture Experiment Base, Changping distract, Beijing (40º10.6'N, 116º26.3'E). The experimental field was about 250 m in length and 80 m in width. The winter wheat was planted in the study area from Oct 3, 2009, and harvested from June 25, 2010. Field inventory was conducted on June 7, 2010 when wheat was in the filling stage. Twenty five ground investigations including different aphid damage levels were selected. Aphid damage level was surveyed according to the investigation rule.

#### **3.1.1.2 Canopy spectral measurements**

Please refer to 1.1.1.2 part above.

### **3.1.2 Results**

62 Remote Sensing – Applications

Fig. 15. Forecast map of yellow rust and distribution of measured points in May, 2008 based

2002; Cooke et al., 2006). The yellow rust is no exception. The weather station can only offer points data, and remote sensing, however, can be a promising means for acquiring spatially continuous observations over large area. It has not been reported, if any, that the LST

The study tried to present a method that could forecast the suitable areas of wheat yellow rust by MODIS temperature products in a large scale. And it was proved that LST derived from remote sensing data had potential for predicting the occurrence and development of wheat yellow rust in a large area. From our results, it is clear that preventive measures of yellow rust can been made over large scale area accordingly with different real-time

Wheat aphid, Sitobion avenae *(Fabricius)*, is one of the most destructive pests in agricultural systems, especially in temperate climates of the northern and southern hemispheres. Wheat aphid appears annually in the wheat planting area of China, causing great economic damage to plant crops as a result of their direct feeding activities. In high enough densities, wheat aphids can remove plant nutrients, and potentially reduce the number of heads, the number of grains per head, and overall seed weight. The damage is especially high when wheat aphid occurs in the flowering and filling stage of wheat. It is reported that average densities over 20 aphids per plant can cause substantial losses of yield and quality of wheat (Basky & Fónagy, 2003). There are also indirect damages including excretion of honeydew from aphids and as a vector of viruses, most notably two strains of the Luteovirus Barley Yellow Dwarf Virus (BYDV-MAV and BYDV-PAV) (Susan et al, 1992). To prevent the

derived from remote sensing data is used to forecast the development of yellow rust.

**3. Detecting and discriminating winter wheat aphid by remote sensing** 

prediction methods based on LST derived from remote sensing data.

on LST

#### **3.1.2.1 Leaf spectral characteristics of wheat infested by aphid**

Representative reflectance measured from wheat aphid-infested and uninfested wheat leaves are shown in Fig. 16. It was evident that the spectral response of the wheat leaf was significantly affected by wheat aphid feeding (Fig. 16). The reflectance of wheat leaf infested by aphid was higher in the visible spectrum and short-wave infrared region and lower in near-infrared region than that of uninfested leaf. A significant increase in the reflectance from the wheat aphid-infested leaf in the visible region (400-700 nm) was observed, evidently due to reduction of photosynthetic pigment concentrations in particular chlorophylls caused by wheat aphid feeding (Richardson et al., 2004).

#### **3.1.2.2 Canopy spectral characteristics of wheat infested by aphid**

Compared with the canopy spectra of the healthy wheat, the canopy reflectance of aphidinfested wheat was gradually decreased in the range from 350 nm to 1750 nm, especially in the near infrared region (Fig. 17). Previous researches indicated that wheat had higher reflectance at visible wavelengths than the healthy vigorously growing wheat because the photoactive pigments (chlorophylls, anthocyanins, carotenoids) were destroyed. In this study, aphid occurred in the filling stage of wheat and the honeydew excreted by aphid absorbed dust or others from surrounding environment and contaminated (darkened) the leaf surface. As a result, the absorption at light slight wavelengths became stronger instead of weaker.

Crop Disease and Pest Monitoring by Remote Sensing 65

visible light, 823 nm (R2=0.865) in near infrared (NIR) and 1654 nm in short-wave infrared

Fig. 18. Correlation coefficient between reflectance and aphid damage levels

ADHI 0.32 0.51

0.17

Aphid damage hyperspectral index (ADHI) was established based on the most sensitive bands from hyperspectral data in the visible light region, NIR and SWIR and weight coefficient calculated according to rate of change of reflectance between healthy wheat and

R551 -R551 R823 - R823

*normal normal*

*infested infested*

*infested*

*normal normal*

R551 R823

*normal*

R1654 - R1654

*normal*

R1654

where R551normal,R823normal and R1654normal are reflectance in 551 nm, 823 nm and 1654 nm of healthy wheat; R551infested , R823 infested , R1654 infested are reflectance in 551 nm, 823 nm and 1654 nm of aphid-infected wheat; 0.32, 0.51 and 0.17 are weight coefficients calculated by the

Further more, the correlation analysis was conducted between ADHI and aphid damage level from 25 investigation points (Fig. 19). It was concluded that ADHI exhibited high relationship with aphid damage levels (R2=0.839*)*. Therefore, ADHI was an important index

Hyperspectral remote sensing has gone through rapid development over the past two decades and there is a trend toward the use of hyperspectral image in the application of remote sensing for precision farming. The study analyzed the spectral characteristics of wheat infested by aphid and selected the sensitive bands to aphid damage level. Then, an ADHI was developed using the most sensitive bands in visible light region, NIR and SWIR.

(SWIR) (R2=0.668), respectively (Fig. 18).

aphid-infected wheat, respectively.

contribution to change rates.

**3.1.3 Conclusions** 

to estimate aphid damage level in winter wheat.

Fig. 16. The spectral reflectance of winter wheat leaf uninfested and infested by aphid

Fig. 17. The spectral reflectance of healthy wheat and wheat infested by various aphid damage levels. (Healthy: the average spectra of healthy wheat samples; Slight: the average spectra of aphid damage level 1and 2; Moderate: he average spectra of aphid damage level 3and 4; Severe: the average spectra of aphid damage level 5 and 6).

#### **3.1.2.3 Aphid damage hyperspectral index for detecting aphid damage degree**

#### **Sensitive band selection of aphid infestation based on canopy reflectance**

The sensitive bands were selected out by relevance analysis between reflectance and aphid damage levels. The reflectance ranges were from 400 nm to 690 nm, from 700 to 1300 nm and from 1500 to 1800 nm. The most sensitive bands to aphid were 551 nm (R2=0.741) in

Fig. 16. The spectral reflectance of winter wheat leaf uninfested and infested by aphid

350 650 950 1250 1550 1850 2150 2450 Wavelength/nm

The sensitive bands were selected out by relevance analysis between reflectance and aphid damage levels. The reflectance ranges were from 400 nm to 690 nm, from 700 to 1300 nm and from 1500 to 1800 nm. The most sensitive bands to aphid were 551 nm (R2=0.741) in

Fig. 17. The spectral reflectance of healthy wheat and wheat infested by various aphid damage levels. (Healthy: the average spectra of healthy wheat samples; Slight: the average spectra of aphid damage level 1and 2; Moderate: he average spectra of aphid damage level

**3.1.2.3 Aphid damage hyperspectral index for detecting aphid damage degree** 

**Sensitive band selection of aphid infestation based on canopy reflectance**

3and 4; Severe: the average spectra of aphid damage level 5 and 6).

Healthy Slight Moderate Severe

0 5

10 15 20

25 30

Reflectance/%

35 40 visible light, 823 nm (R2=0.865) in near infrared (NIR) and 1654 nm in short-wave infrared (SWIR) (R2=0.668), respectively (Fig. 18).

Fig. 18. Correlation coefficient between reflectance and aphid damage levels

Aphid damage hyperspectral index (ADHI) was established based on the most sensitive bands from hyperspectral data in the visible light region, NIR and SWIR and weight coefficient calculated according to rate of change of reflectance between healthy wheat and aphid-infected wheat, respectively.

$$\text{ADHI} = 0.32 \times \frac{\text{R551}\_{normal} \quad \text{-R551}\_{infeeded}}{\text{R551}\_{normal}} + 0.51 \times \frac{\text{R823}\_{normal} \quad \text{- R823}\_{infeeded}}{\text{R823}\_{normal}}$$

$$+0.17 \times \frac{\text{R1654}\_{normal} \quad \text{- R1654}\_{infeeded}}{\text{R1654}\_{normal}}$$

where R551normal,R823normal and R1654normal are reflectance in 551 nm, 823 nm and 1654 nm of healthy wheat; R551infested , R823 infested , R1654 infested are reflectance in 551 nm, 823 nm and 1654 nm of aphid-infected wheat; 0.32, 0.51 and 0.17 are weight coefficients calculated by the contribution to change rates.

Further more, the correlation analysis was conducted between ADHI and aphid damage level from 25 investigation points (Fig. 19). It was concluded that ADHI exhibited high relationship with aphid damage levels (R2=0.839*)*. Therefore, ADHI was an important index to estimate aphid damage level in winter wheat.

#### **3.1.3 Conclusions**

Hyperspectral remote sensing has gone through rapid development over the past two decades and there is a trend toward the use of hyperspectral image in the application of remote sensing for precision farming. The study analyzed the spectral characteristics of wheat infested by aphid and selected the sensitive bands to aphid damage level. Then, an ADHI was developed using the most sensitive bands in visible light region, NIR and SWIR.

Crop Disease and Pest Monitoring by Remote Sensing 67

Fig. 20. The study area and the spatial distribution of sample plots

year.

study areas have flat topography, with elevation ranging from 20 m to 40 m. The study areas have semi-humid warm temperate climate with yearly precipitation of 625 mm and mean temperature of 11.5°C in the Shunyi district and yearly precipitation 620 mm and mean temperature of 11.3°C in the Tongzhou district. Both districts are considered main winter wheat planting areas in Beijing, and aphid infestations occur in both areas almost every

a) b)

Fig. 19. The correlation between ADHI and aphid damage level

It was concluded that ADHI was a sensitive index to aphid damage levels, and could be used to retrieve aphid damage levels in the filling stage of wheat.

Crop growth is very dynamic processes and monitoring the condition of agricultural corps is a complex issue. It is possible that wheat damage symptoms caused by aphids and its response of canopy reflectance are different in different wheat growth stages. This study revealed that the reflectance of wheat infested by aphid was lower than healthy wheat in filling stage probably because of honeydew excreted by aphid. This was not consistent with previously published results in early detection of aphid infestation. Therefore, whether the ADHI can effectively retrieve aphid damage levels in other wheat growth stages remains as a task of future studies.

#### **3.2 Detecting winter wheat aphid incidence using Landsat 5 TM**

Wheat aphid occurrence and damage degrees are related to many factors including temperature, humidity, precipitation, field management, enemies, etc.. Most of the present studies on aphid prediction have been conducted based on meteorological data acquired from weather stations, and aphid density was monitored using the spectral characteristics of wheat infested by aphid. However, it is rare to investigate the relationship between environmental parameters, vegetable information derived from satellite images and aphid damage degrees. The aim of the present study is to investigate the relationships of aphid occurrence and damage degree to LST, NDWI, and MNDWI, which are related to vegetation water content derived from multi-temporal Landsat 5 TM. Another goal of the current research is to distinguish the degrees of aphid damage using 2-dimension feature spaces established by LST-NDWI and LST-MNDWI.

#### **3.2.1 Materials and methods**

#### **3.2.1.1 Study areas**

The study areas are selected in Shunyi district (116°28'—116°58' E,40°00'—40°18' N) and Tongzhou district (116° 32'—116°56' E, 39°36' —40°02' N,) of Beijing, China (Fig.20-a). The

R2 = 0.839


It was concluded that ADHI was a sensitive index to aphid damage levels, and could be

Crop growth is very dynamic processes and monitoring the condition of agricultural corps is a complex issue. It is possible that wheat damage symptoms caused by aphids and its response of canopy reflectance are different in different wheat growth stages. This study revealed that the reflectance of wheat infested by aphid was lower than healthy wheat in filling stage probably because of honeydew excreted by aphid. This was not consistent with previously published results in early detection of aphid infestation. Therefore, whether the ADHI can effectively retrieve aphid damage levels in other wheat growth stages remains as

Wheat aphid occurrence and damage degrees are related to many factors including temperature, humidity, precipitation, field management, enemies, etc.. Most of the present studies on aphid prediction have been conducted based on meteorological data acquired from weather stations, and aphid density was monitored using the spectral characteristics of wheat infested by aphid. However, it is rare to investigate the relationship between environmental parameters, vegetable information derived from satellite images and aphid damage degrees. The aim of the present study is to investigate the relationships of aphid occurrence and damage degree to LST, NDWI, and MNDWI, which are related to vegetation water content derived from multi-temporal Landsat 5 TM. Another goal of the current research is to distinguish the degrees of aphid damage using 2-dimension feature spaces

The study areas are selected in Shunyi district (116°28'—116°58' E,40°00'—40°18' N) and Tongzhou district (116° 32'—116°56' E, 39°36' —40°02' N,) of Beijing, China (Fig.20-a). The

0

a task of future studies.

Fig. 19. The correlation between ADHI and aphid damage level

used to retrieve aphid damage levels in the filling stage of wheat.

**3.2 Detecting winter wheat aphid incidence using Landsat 5 TM** 

established by LST-NDWI and LST-MNDWI.

**3.2.1 Materials and methods** 

**3.2.1.1 Study areas** 

1

2

3

4

Aphid damage level

5

6

7

Fig. 20. The study area and the spatial distribution of sample plots

study areas have flat topography, with elevation ranging from 20 m to 40 m. The study areas have semi-humid warm temperate climate with yearly precipitation of 625 mm and mean temperature of 11.5°C in the Shunyi district and yearly precipitation 620 mm and mean temperature of 11.3°C in the Tongzhou district. Both districts are considered main winter wheat planting areas in Beijing, and aphid infestations occur in both areas almost every year.

Crop Disease and Pest Monitoring by Remote Sensing 69

LST is the radioactive skin temperature of the land surface, which plays an important role in farm and ecological environment. The present paper aims to discuss the relationship between LST and aphid occurrence and spread. LST was derived from the thermal infrared band (10.4-12.5μm) data of Landsat-5 TM using generalized single-channel algorithm developed by Jiménez-Muñoz and Sobrino (Jiménez-Muñoz and Sobrino, 2004). Surface emissivity (ε) and atmospheric water vapor content (w) were important parameters in the generalized single-channel algorithm. In the study, w was derived from the reflectance of band2 and band19 of MOD02, (Kaufman and Gao, 1992), and ε was calculated by vegetation

The NDWI, MNDWI and LST of all sample points were calculated and extracted from the

We resized the subset areas with size of 7.2 km2 (3 km × 2.4 km) from the study area image located in Tongzhou district and covered with 20 evenly distributed sample points, and the aphid densities of the sample points were surveyed on May 6, May 20 and June 4, 2010, respectively. The survey results showed that the aphid damage degree of all sample points were S0 on May 6, 18 points for S1 and 2 points for S0 on May 20, and 16 points for S2 and 4 points for S0 on June 4, respectively. The subset areas were small enough and 20 sample points evenly distributed, According to the survey result, the aphid damage degree of the sample plots was basically same. Thus, the change of the aphid damage degree of wheat pixels in the wheat plots was slim or even basically the same as the sample plots. The wheat area of subset image selection area was extracted using classification of decision tree in ENVI 4.5 (Fig 20-b). The LST, NDWI and MNDWI of 2000 wheat pixels were extracted.

One basic accuracy assessment currently being used is overall accuracy, which is calculated by dividing the correctly classified pixels by the total number of the pixels checked. The Kappa coefficient is a measure of the overall agreement of a matrix introduced to the remote sensing community in early 1983. It has since become a widely used measure for classification accuracy. In contrast to overall accuracy, the Kappa coefficient takes non-diagonal elements into account (Rosenfield and Fitzpatrick-Lins, 1986), and it is calculated by the formula:

r r

1 1 <sup>r</sup> <sup>2</sup>

*NX XX*

*i*

where r is the number of rows and columns in the error matrix; N is the total number of observations; Xii is the observation in row i and column i; Xi+ is the marginal total of row I;

The minimum value, maximum value, mean values and standard deviations of LST, NDWI and MNDWI with aphid damage degrees of wheat pixels in subset image selection were

*i i*

*K*

1

*N XX*

*ii i i*

*i i*

coverage (Carlson and Ripley, 1997).

**3.2.1.6 Methods of accuracy assessment** 

X+i is the marginal total of column i.

**3.2.2.1 2-dimensional feature space based on LST-VI** 

**3.2.2 Results** 

**3.2.1.5 Subset image selection and wheat extraction** 

Landsat images.

#### **3.2.1.2 Field inventory and data pre-processing**

Field inventory was conducted during the growing seasons of winter wheat in 2010. The winter wheat in the study areas were planted between September 25 and October 7, 2009, and harvested between June 19 and June 25, 2010. Based on the combination of representative sampling and random sampling scheme, 70 sample plots with size of 0.09 ha (30 m × 30 m) each were collected as in Fig 1-a. These sample plots had different site conditions, plant densities, and management conditions. Aphid density surveys were carried out respectively on May 4 and May 6 for jointing stage, May 20 and May 21 for the heading stage, and June 3 and June 4 for the filling stage. The geographical coordinates of each plot were measured by global positioning system (GPS) ( GeoExplorer 3000 GPS, with the error within 1m) at the middlemost of the plot.

Each sample covered with an area of 1 m2. Then, 10 tillers in each sample plot were randomly selected, and the number of aphids was counted. The aphid densities were then estimated as follows: total aphids /10 tillers.

The survey results were divided into three aphid damage degrees according to the aphid density investigated for facilitating the study. They were S0: non-infested by aphid and no damage to wheat, S1: aphid abundance/per tiller was about 3-10 and damage degree to wheat was slight, and S3: aphid abundance/per tiller was more than 20 and damage degree to wheat was severe.

#### **3.2.1.3 Satellite image acquisition and pre-processing**

Three Landsat-5 Thematic Mapper (TM) images (path 123/row 32) and three MOD 02 1 KM-Level 1B Calibrated Radiances Production (MOD 02) were acquired on May 4, May 20 and June 5, 2010, respectively. And all images were more than 90% cloud-free.

The Landsat-5 TM images were spectrally corrected to reflectance using the Landsat TM calibration tool and FLAASH (Fast line-of-sight Atmospherics Analysis of Spectral Hypercubes) was used to correct the image for atmospheric effects in ENVI 4.5. The Landsat-5 TM images were geometrically corrected versus a reference IKONOS image (equivalent scale map 1:10000) of the same area, available from a previous study. The resulting root mean square error (RMSE) did not exceed 0.3 pixels, which was adequate for the purposes of the present study.

#### **3.2.1.4 Derivation of LST, NDWI and MNDWI from Landsat 5 TM**

NDWI and MNDWI are both sensitive to changes in liquid water content of vegetation canopies (Hunt and Rock, 1989). In the current research, both NDWI and MNDWI were used to determine the threshold of aphid occurrence and the aphid damage degree. The indices are of the general form, as shown in the following:

$$\text{NDDVVI} = \frac{\text{R}\_{\text{NIR}} - \text{R}\_{\text{SWIR}}}{\text{R}\_{\text{NIR}} + \text{R}\_{\text{SWIR}}} \quad \text{MINIDVI} = \frac{\text{R}\_{\text{GREEN}} \cdot \text{R}\_{\text{SWIR}}}{\text{R}\_{\text{GREEN}} + \text{R}\_{\text{SWIR}}}$$

where R*GREEN* ,R*NIR* and R*SWIR* are the reflectance in the green band, near-infrared band and short wave infrared band, respectively. For Landsat TM/ETM+, R*GREEN* ,R*NIR* and R*SWIR*  correspond to band2, band4 and band5, respectively.

Field inventory was conducted during the growing seasons of winter wheat in 2010. The winter wheat in the study areas were planted between September 25 and October 7, 2009, and harvested between June 19 and June 25, 2010. Based on the combination of representative sampling and random sampling scheme, 70 sample plots with size of 0.09 ha (30 m × 30 m) each were collected as in Fig 1-a. These sample plots had different site conditions, plant densities, and management conditions. Aphid density surveys were carried out respectively on May 4 and May 6 for jointing stage, May 20 and May 21 for the heading stage, and June 3 and June 4 for the filling stage. The geographical coordinates of each plot were measured by global positioning system (GPS) ( GeoExplorer 3000 GPS, with

Each sample covered with an area of 1 m2. Then, 10 tillers in each sample plot were randomly selected, and the number of aphids was counted. The aphid densities were then

The survey results were divided into three aphid damage degrees according to the aphid density investigated for facilitating the study. They were S0: non-infested by aphid and no damage to wheat, S1: aphid abundance/per tiller was about 3-10 and damage degree to wheat was slight, and S3: aphid abundance/per tiller was more than 20 and damage degree

Three Landsat-5 Thematic Mapper (TM) images (path 123/row 32) and three MOD 02 1 KM-Level 1B Calibrated Radiances Production (MOD 02) were acquired on May 4, May 20

The Landsat-5 TM images were spectrally corrected to reflectance using the Landsat TM calibration tool and FLAASH (Fast line-of-sight Atmospherics Analysis of Spectral Hypercubes) was used to correct the image for atmospheric effects in ENVI 4.5. The Landsat-5 TM images were geometrically corrected versus a reference IKONOS image (equivalent scale map 1:10000) of the same area, available from a previous study. The resulting root mean square error (RMSE) did not exceed 0.3 pixels, which was adequate for

NDWI and MNDWI are both sensitive to changes in liquid water content of vegetation canopies (Hunt and Rock, 1989). In the current research, both NDWI and MNDWI were used to determine the threshold of aphid occurrence and the aphid damage degree. The


where R*GREEN* ,R*NIR* and R*SWIR* are the reflectance in the green band, near-infrared band and short wave infrared band, respectively. For Landsat TM/ETM+, R*GREEN* ,R*NIR* and R*SWIR* 

*R R MNDWI*

+ *GREEN SWIR GREEN SWIR*

*R R*

and June 5, 2010, respectively. And all images were more than 90% cloud-free.

**3.2.1.4 Derivation of LST, NDWI and MNDWI from Landsat 5 TM** 

+ *NIR SWIR NIR SWIR*

*R R*

indices are of the general form, as shown in the following:

*R R NDWI*

correspond to band2, band4 and band5, respectively.

**3.2.1.2 Field inventory and data pre-processing** 

the error within 1m) at the middlemost of the plot.

**3.2.1.3 Satellite image acquisition and pre-processing** 

estimated as follows: total aphids /10 tillers.

to wheat was severe.

the purposes of the present study.

LST is the radioactive skin temperature of the land surface, which plays an important role in farm and ecological environment. The present paper aims to discuss the relationship between LST and aphid occurrence and spread. LST was derived from the thermal infrared band (10.4-12.5μm) data of Landsat-5 TM using generalized single-channel algorithm developed by Jiménez-Muñoz and Sobrino (Jiménez-Muñoz and Sobrino, 2004). Surface emissivity (ε) and atmospheric water vapor content (w) were important parameters in the generalized single-channel algorithm. In the study, w was derived from the reflectance of band2 and band19 of MOD02, (Kaufman and Gao, 1992), and ε was calculated by vegetation coverage (Carlson and Ripley, 1997).

The NDWI, MNDWI and LST of all sample points were calculated and extracted from the Landsat images.

#### **3.2.1.5 Subset image selection and wheat extraction**

We resized the subset areas with size of 7.2 km2 (3 km × 2.4 km) from the study area image located in Tongzhou district and covered with 20 evenly distributed sample points, and the aphid densities of the sample points were surveyed on May 6, May 20 and June 4, 2010, respectively. The survey results showed that the aphid damage degree of all sample points were S0 on May 6, 18 points for S1 and 2 points for S0 on May 20, and 16 points for S2 and 4 points for S0 on June 4, respectively. The subset areas were small enough and 20 sample points evenly distributed, According to the survey result, the aphid damage degree of the sample plots was basically same. Thus, the change of the aphid damage degree of wheat pixels in the wheat plots was slim or even basically the same as the sample plots. The wheat area of subset image selection area was extracted using classification of decision tree in ENVI 4.5 (Fig 20-b). The LST, NDWI and MNDWI of 2000 wheat pixels were extracted.

#### **3.2.1.6 Methods of accuracy assessment**

One basic accuracy assessment currently being used is overall accuracy, which is calculated by dividing the correctly classified pixels by the total number of the pixels checked. The Kappa coefficient is a measure of the overall agreement of a matrix introduced to the remote sensing community in early 1983. It has since become a widely used measure for classification accuracy. In contrast to overall accuracy, the Kappa coefficient takes non-diagonal elements into account (Rosenfield and Fitzpatrick-Lins, 1986), and it is calculated by the formula:

$$K = \frac{N\sum\_{i=1}^{\mathbf{r}} X\_{ii} - \sum\_{i=1}^{\mathbf{r}} X\_i + X\_{+i}}{N^2 - \sum\_{i=1}^{\mathbf{r}} X\_i + X\_{+i}}$$

where r is the number of rows and columns in the error matrix; N is the total number of observations; Xii is the observation in row i and column i; Xi+ is the marginal total of row I; X+i is the marginal total of column i.

#### **3.2.2 Results**

#### **3.2.2.1 2-dimensional feature space based on LST-VI**

The minimum value, maximum value, mean values and standard deviations of LST, NDWI and MNDWI with aphid damage degrees of wheat pixels in subset image selection were

Crop Disease and Pest Monitoring by Remote Sensing 71

In the 2-dimensional feature space coordinate system that was composed by LST and MNDWI, the S0 samples mainly scattered on the left part of the coordinate system, whereas S1 and S2 samples were distributed on the right part. As shown in Fig. 22, when LST was lower than the certain value, aphid did not occur, suggesting that LST served as a key factor

Furthermore, LST0 and MNDWI0, which were the cutoff value of threshold values of LST and MNDWI of S0, S1 and S2, were determined by mean values and standard deviations.

LST0 =LST\_M1-2×LST\_SD1

MNDWI0= (M\_M1+3×M\_SD1)+ [(M\_M1+ 3×M\_SD1)-(M\_M2-3×M\_SD2)]/2 where LST\_M1 and LST\_SD1 are the mean value and standard deviation of LST for S1; M\_M1and M\_SD1 are the mean value and standard deviation of MNDWI for S1; and M\_M2

According to Table 3, LST0 = 297.7568 and MNDWI0 = -0.3866. Wheat was not infested by aphid when LST< 297.7568, and aphid damage degree was S1 when LST≥297.7568K and - 0.6506≤MDNWI ≤-0.3866 and S2 when LST≥297.7568K and -0.3866 ≤MDNWI ≤-0.1077 (Fig.

**3.2.2.2 Discriminating aphid damage degrees using LST and MNDWI** 

of aphid occurrence and the MNDWI was sensitive to aphid damage degree.

and M\_SD2 are the mean value and standard deviation of MNDWI for S2.

Fig. 22. Discriminating aphid damage degrees using LST and MNDWI

All survey samples, except 20 samples in the subset selection image were used to test the aphid prediction accuracy of 2-dimensional feature space based on LST and MNDWI (Fig.

The discrimination accuracy was assessed using overall accuracy and kappa coefficient (Table 11). The results showed that the overall accuracy was 84%, and the Kappa accuracy

LST0 and MNDWI0 were calculated by formula as follows:

22).

**3.2.2.3 Verification** 

23).

was 75.67%.

listed in Table 9 and Table 10. And 2-dimensional feature space coordinates were established with LST as the abscissa and NDWI and MNDWI as the vertical axis, respectively (Figs. 2, 3). LST ranged from 287.5879 to 313.3448, NDWI ranged from 0.0226 to 0.5591 and MNDWI ranged from -0.3402 to -0.1077, respectively.

It is clear that LST was increasing from S0 to S1 to S2. LST was an important driving factor for aphid occurrence and could distinguish wheat non-infected from infested by aphids (Fig. 21 and Table 9). The general trend of NDWI increased firstly and reduced afterward, whereas MNDWI reduced firstly and increased afterward from S0 to S1 to S2.


Table 9. Minimum and maximum values of LST, NDWI and MNDWI in S0, S1 and S2


Table 10. Mean value and standard derivation of LST, NDWI and MNDWI in S0, S1 and S2

Fig. 21. The distribution of S0, S1 and S2 in the LST-NDWI (left) and LST-MNDWI (right) feature space

listed in Table 9 and Table 10. And 2-dimensional feature space coordinates were established with LST as the abscissa and NDWI and MNDWI as the vertical axis, respectively (Figs. 2, 3). LST ranged from 287.5879 to 313.3448, NDWI ranged from 0.0226 to

It is clear that LST was increasing from S0 to S1 to S2. LST was an important driving factor for aphid occurrence and could distinguish wheat non-infected from infested by aphids (Fig. 21 and Table 9). The general trend of NDWI increased firstly and reduced afterward,

> Minimum value

Table 9. Minimum and maximum values of LST, NDWI and MNDWI in S0, S1 and S2

LST NDWI MNDWI

Mean value

S0 290.8578 1.4740 0.3029 0.0574 -0.2293 0.0296 S1 299.9236 1.0834 0.3998 0.0587 -0.4940 0.0362 S2 303.9424 1.7121 0.2979 0.0458 -0.2672 0.0402

Table 10. Mean value and standard derivation of LST, NDWI and MNDWI in S0, S1 and S2

Fig. 21. The distribution of S0, S1 and S2 in the LST-NDWI (left) and LST-MNDWI (right)

S0 287.5879 296.2498 0.0226 0.4405 -0.3402 -0.1077 S1 297.8084 306.0133 0.2083 0.5591 -0.6506 -0.3326 S2 300.5391 313.3448 0.0473 0.4542 -0.4117 -0.1159

LST NDWI MNDWI

Standard

Maximum value

Minimum value

deviation Mean value Standard

Maximum value

deviation

0.5591 and MNDWI ranged from -0.3402 to -0.1077, respectively.

Maximum value

Aphid Damage Degree

Aphid Damage Degree

feature space

Minimum value

Mean value Standard

deviation

whereas MNDWI reduced firstly and increased afterward from S0 to S1 to S2.

#### **3.2.2.2 Discriminating aphid damage degrees using LST and MNDWI**

In the 2-dimensional feature space coordinate system that was composed by LST and MNDWI, the S0 samples mainly scattered on the left part of the coordinate system, whereas S1 and S2 samples were distributed on the right part. As shown in Fig. 22, when LST was lower than the certain value, aphid did not occur, suggesting that LST served as a key factor of aphid occurrence and the MNDWI was sensitive to aphid damage degree.

Furthermore, LST0 and MNDWI0, which were the cutoff value of threshold values of LST and MNDWI of S0, S1 and S2, were determined by mean values and standard deviations. LST0 and MNDWI0 were calculated by formula as follows:

LST0 =LST\_M1-2×LST\_SD1

MNDWI0= (M\_M1+3×M\_SD1)+ [(M\_M1+ 3×M\_SD1)-(M\_M2-3×M\_SD2)]/2

where LST\_M1 and LST\_SD1 are the mean value and standard deviation of LST for S1; M\_M1and M\_SD1 are the mean value and standard deviation of MNDWI for S1; and M\_M2 and M\_SD2 are the mean value and standard deviation of MNDWI for S2.

According to Table 3, LST0 = 297.7568 and MNDWI0 = -0.3866. Wheat was not infested by aphid when LST< 297.7568, and aphid damage degree was S1 when LST≥297.7568K and - 0.6506≤MDNWI ≤-0.3866 and S2 when LST≥297.7568K and -0.3866 ≤MDNWI ≤-0.1077 (Fig. 22).

Fig. 22. Discriminating aphid damage degrees using LST and MNDWI

#### **3.2.2.3 Verification**

All survey samples, except 20 samples in the subset selection image were used to test the aphid prediction accuracy of 2-dimensional feature space based on LST and MNDWI (Fig. 23).

The discrimination accuracy was assessed using overall accuracy and kappa coefficient (Table 11). The results showed that the overall accuracy was 84%, and the Kappa accuracy was 75.67%.

Crop Disease and Pest Monitoring by Remote Sensing 73

Becker, B. L., David, P. L., & Qi, J. G. (2007). A classification-based assessment of the optimal

Bravo, C., Moshou, D., West, J., McCartney, A., & Ramon, H. (2003). Early disease detection in wheat elds using spectral reectance. *Biosystems Engineering*, 84, 137–145. Broge, N. H., & E. Leblanc. 2000. Comparing prediction power and stability of broadband

canopy chlorophyll density. *Remote Sensing of Environment,* 76, 156–72. Carlson T. N. & Ripley D. A.. On the relation between NDVI, fractional vegetation cover, and leaf area index. *Remote Sensing of Environment,* 1997, 62(3): 241-252. Ceccato, P., N. Gobron, S. Flasse, B. Pinty, & S. Tarantola. 2002. Designing a spectral index to

approach. *Remote Sensing of Environment*, 82, 188–97.

food insecurity. *Nutrition Research Reviews* 17, 23–42.

*Remote Sensing of Environment*, 74, 229-239.

remote sensing. *Precision Agriculture*, 8, 161–172.

*Sensing of Environment,* 87, 111–21.

523–34.

41(1), 35–44.

416–426.

*and Photobiology*, 74, 38–45.

*Environment*, 95, 351–367.

*Sensing of Environment*, 108, 111–120

spectral and spatial resolutions for Great Lakes coastal wetland imagery. *Remote* 

and hyperspectral vegetation indices for estimation of green leaf area index and

estimate vegetation water content from remote sensing data: Part 1. Theoretical

Estimating corn leaf chlorophyll concentration from leaf and canopy reflectance.

from MODIS near-and shortwave infrared data in a semiarid environment. *Remote* 

Southeastern Brazil with EO-1 Hyperion data. *Remote Sensing of Environment*, 94,

tracks diurnal changes in photosynthetic efciency. *Remote Sensing of Environment*,

nondestructive estimation of anthocyanin content in plant leaves. *Photochemistry* 

high resolution hyperspectral imagery for detection of anomalies. *Remote Sensing of* 

Integrated narrowband vegetation indices for prediction of crop chlorophyll content for application to precision agriculture. *Remote Sensing of Environment*, 81,

Hyperspectral vegetation indices and novel algorithms for predicting green LAI of

Christou, P., Twyman, R.M., (2004). The potential of genetically enhanced plants to address

Daughtry, C. S., C. L. Walthall, M. S. Kim, E. Brown de Colstoun, & J. E. McMurtrey.(2000).

Fensholt, R., & I. Sandholt.(2003). Derivation of a shortwave infrared water stress index

Filella, I., Serrano, L., Serra, J., & Penuelas, J. (1995). Evaluating wheat nitrogen status with canopyreectance indices and discriminant analysis. *Crop Science*, 35, 1400–1405. Franke, J., & Menz, G. (2007). Multi-temporal wheat disease detection by multi-spectral

Galvão, L. S., A. Formaggio, R. & Tisot, D. A. (2005). Discrimination of sugarcane varieties in

Gamon, J. A., Penuelas, J., & Field, C. B. (1992). A narrow-waveband spectral index that

Gitelson, A. A., Merzlyak, M. N., & Chivkunova, O. B. (2001). Optical properties and

Goovaerts, P., Jacquez, G. M., & Marcus, A. (2005). Geostatistical and local cluster analysis of

Gong, P., Pu R., Heald R.C.(2002). Analysis of in situ hyperspectral data for nutrient estimation of giant sequoia. *International Journal of Remote Sensing,* 23, 1827-1850. Haboudane, D., Miller, J. R., Tremblay, N., Zarco-Tejada, P. J., & Dextraze, L. (2002).

Haboudane, D., J. R. Miller, E. Pattery, P. J. Zarco-Tejad, & I. B. Strachan. (2004).

Fig. 23. Distribution of test sample plots in LST-MNDWI feature space


Kappa coefficient = 0.7567

Table 11. Error matrices of the verification samples

#### **3.2.3 Conclusions**

This study successfully investigated the relationship between aphid damage degrees and several spectral features, such as NDWI, MNDWI and LST, through 2-dimensional feature space method. The results indicated that LST was the key factor in predicting the occurrence of aphid, and MNDWI was more sensitive to aphid damage degree than NDWI. In the 2 dimension feather space composed by LST and MNDWI, the result showed that S0, S1 and S2 were divided into three regions; S0 was distributed on the left of the space, and S1 and S2 on the right. Further, LST0 and MNDWI0 were calculated according the mean and derivation of S1, S2 as the cutoff value of threshold value to discriminate S0, S1 and S0. Through the verification of discrimination threshold value, it confirmed that the overall accuracy of discrimination was 84% and Kappa coefficient was 0.7567, suggesting that LST and MNDWI were of great potential in discriminating and monitoring the aphid damage degree over a large area, only using thermal infrared band and multi-spectral satellite images.

#### **4. References**

Basky Z. & Fónagy A. (2003).Glutenin and gliadin contents of flour derived from wheat infested with different aphid species. *Pest Management Science*, 59, 426-430.

 S0 S1 S2 Total S0 17 0 0 17 S1 2 14 0 16 S2 4 2 11 19 Total 23 16 11 50

This study successfully investigated the relationship between aphid damage degrees and several spectral features, such as NDWI, MNDWI and LST, through 2-dimensional feature space method. The results indicated that LST was the key factor in predicting the occurrence of aphid, and MNDWI was more sensitive to aphid damage degree than NDWI. In the 2 dimension feather space composed by LST and MNDWI, the result showed that S0, S1 and S2 were divided into three regions; S0 was distributed on the left of the space, and S1 and S2 on the right. Further, LST0 and MNDWI0 were calculated according the mean and derivation of S1, S2 as the cutoff value of threshold value to discriminate S0, S1 and S0. Through the verification of discrimination threshold value, it confirmed that the overall accuracy of discrimination was 84% and Kappa coefficient was 0.7567, suggesting that LST and MNDWI were of great potential in discriminating and monitoring the aphid damage degree over a large area, only using thermal infrared band and multi-spectral satellite

Basky Z. & Fónagy A. (2003).Glutenin and gliadin contents of flour derived from wheat infested with different aphid species. *Pest Management Science*, 59, 426-430.

Fig. 23. Distribution of test sample plots in LST-MNDWI feature space

Kappa coefficient = 0.7567

**3.2.3 Conclusions** 

images.

**4. References** 

Table 11. Error matrices of the verification samples


Crop Disease and Pest Monitoring by Remote Sensing 75

Peñuelas, J., J. Piñol, R. Ogaya, & I. Filella. (1997). Estimation of plant water concentration by

Pu, R., Ge S., Kelly N.M., Gong P. (2003). Spectral absorption features as indicators of water

Pu, R., Foschi L., Gong P. (2004). Spectral feature analysis for assessment of water status and

Rouse, J. W., R. H. Haas, J. A. Schell, & D. W. Deering.(1973). Monitoring vegetation systems

Rosenfield G. & Fitzpatrick-Lins K. (1986). A coefficient of agreement as a measure of

Rules for Resistance Evaluation of Wheat to Diseases and Insect Pests Part 7: Rule for

Raikes, C., & L. L. Burpee. (1998). Use of multispectral radiometry for assessment of Rhizoctonia blight in creeping bentgrass. *Phytopathology*, 88, 446-449. South, S., Qi, J. G., & Lusch, D. P. (2004). Optimal classification methods for mapping agricultural tillage practices. *Remote Sensing of Environment*, 91, 90–97 Susan E. Halbert, June Connelly B., Bishop G. W., et al. (1992).Transmission of barley yellow

Strange, R.N., Scott, P.R., (2005). Plant Disease: A Threat to Global Food Security. *Annual* 

Sutton, J.C., Gillespie, T.J., Hildebrand, P.D. (1984). Monitoring weather factors in relation to

Thenkabail, P. S., Smith, R. B., & De Pauw, E. (2000). Hyperspectral vegetation indices and

West, J. S., Bravo, C., Oberti, R., Lemaire, D., Moshou, D., & McCartney, H. A. (2003). The

Xiao, Z. Q., Li, Z. M., Fan M., Zhang, Y. Ma, S. j. (2007), Prediction model on stripe rust

Yu, B., Ostland, I.M., Gong, P., Pu, R. L. (1999) Penalized discriminant analysis of in situ

Zarco-Tejada, P. J., A. Berjón, R. López-Lozano, J. R. Miller, P. Martín, V. Cachorro, M. R.

Zeng, S. M. (2003). Simulation study on oversummering process of wheat stripe rust caused Puccinia striiformis west. *In China, Acta Phytopathologica Sinica*, 33, 267-278.

in the Great Plains with ERTS. *Proc 3rd ERTS Symp* 1, 48–62.

Resistance Evaluation of Wheat to Aphids. NY/T 1443.7-2007

18, 2869–75.

223-227.

*Sensing*, 24, 1799-1810.

*Remote Sensing*, 25, 4267-4286.

*Applied Biology*, 121, 105-121.

*Environment*, 71, 158–182.

*Agrometeorology*, 28, 350-353.

*and remote sensing*, 37, 2569-2577.

*review of Phytopathology*, 40, 83–116.

plant disease. *Plant Disease*, 68, 78–84.

*Annual Reviews of Phytopathology*, 41, 593–614.

canopy. *Remote Sensing Environment*, 99:271–87.

the reflectance water index WI (R900/R970). *International Journal of Remote Sensing,*

status in coast live oak (Quercus agrifolia) leaves. *International Journal of Remote* 

health level in coast live oak (Quercus agrifolia) leaves. *International Journal of* 

thematic classification accuracy. *Photogrammetric Engineering and Remote Sensing,* 52,

dwarf virus by field collected aphids (Homoptera: Aphididae) and their relative importance in barley yellow dwarf epidemiology in southwestern Idaho. *Annals of* 

their relationships with agricultural crop characteristics. *Remote Sensing of* 

potential of optical canopy measurement for targeted control of eld crop disease.

influence extent of winter wheat in Longnan Mountain area. *Chinese Journal of* 

hyperspectral data for conifer species recognition. *IEEE Transactions on geosciences* 

González, & A. Frutos. (2005). Assessing vineyard condition with hyperspectral indices: Leaf and canopy reflectance simulation in a row-structured discontinuous

crop canopies: Modeling and validation in the context of precision agriculture. *Remote Sensing of Environment,* 90, 337–52.


Hélène L., Frédéric F., Pierre D., Gaétan B., Isztar Z. (2002),Estimation of the spatial pattern

Huang, W. J., David, W. L., Niu, Z., Zhang, Y. J., Liu, L. Y., & Wang, J. H. (2007).

Hunt E.R., Rock B.N. (1989). Detection of changes in leaf water content using Near- and Middle-Infrared reflectances. *Remote Sensing of Environment*, 30, 43-54. Jiménez-Muñoz J. C., Sobrino J. A. & Leonardo P. (2004). Land Surface Temperature Retrieval from Landsat TM 5. *Remote Sensing of Environment*, 90, 434-440. Kaufman Y. J. & Gao B. C. (1992). Remote Sensing of Water Vapor in the Near IR from EOS/MODIS. *IEEE Transactions on Geoscience and Remote Sensing*, 30, 871-884. Kim, M. S., C. S. T. Daughtry, E. W. Chappelle, & J. E. McMurtrey. 1994. The use of high

Li, G. B., Zeng, S. M., & Li, Z. Q. (1989). Integrated Management of Wheat Pests (pp. 185– 186). Beijing: Press of Agriculture Science and Technology of China (in Chinese). Luo, J. H., Zhang, J. C., Huang, W. J., Xu, X. G., Jin, N. (2010). Preliminary study on the

Luo, J. H., Huang, W. J., Zhang, J. C., Xu X. G. & Wang D. C. (2011). The preliminary study

Merton, R., & J. Huntington.(1999). Early simulation of the ARIES-1 satellite sensor for

Moshou, D., Bravo, C., West, J., Wahlen, S., McCartney, A., & Ramon, H. (2004). Automatic

Peñuelas, J., Baret, F., & Filella, I.(1995). Semi-empirical indices to assess

Merzlyak, M. N., Gitelson, A. A., Chivkunova, O. B., & Rakitin, V. Y. (1999). Non-destructive

Peñuelas, J., Gamon, J. A., Fredeen, A. L., Merino, J., & Field, C. B. (1994). Reectance indices

networks. *Computers and Electronics in Agricultur*e, 44, 173–188.

*Measurements and Signatures in Remote Sensing*, 299–306.

winter wheat. *Disaster Advances,* 3, 288-292.

remote sensing. *Sensor Letters,* 9, 1225-1228.

*Physiologia Plantarum*, 106, 135–141.

leaves. *Remote Sensing of Environment*, 48, 135–146.

Publication 99-17.

221–230.

*Remote Sensing of Environment,* 90, 337–52.

231.

197.

crop canopies: Modeling and validation in the context of precision agriculture.

of surface relative humidity using ground based radar measurements and its application to disease risk assessment. *Agricultural and Forest Meteorology*, 111, 223-

Identification of yellow rust in wheat using in situ spectral reflectance measurements and airborne hyperspectral imaging. *Precision Agriculture*, 8, 187–

spectral resolution bands for estimating absorbed photosynthetically active radiation (APAR). *In Proceedings of the 6th International Symposium on Physical* 

relationship between land surface temperature and occurrence of yellow rust in

on spectral response of wheat under different stresses between field and satellite

multi-temporal vegetation research derived from AVIRIS. In *Summaries of the Eight JPL Airborne Earth Science Workshop*, 9–11 February, 299–307. Pasadena, CA: JPL

detection of'yellow rust' in wheat using reflectance measurements and neural

carotenoids/chlorophyll a ratio from leaf spectral reflectance. *Photosynthetica*, 31,

optical detection of pigment changes during leaf senescence and fruit ripening.

associated with physiological changes in nitrogen- and water-limited sunower


**3** 

*Brazil* 

**Seasonal Variability of Vegetation** 

*Brazilian National Institute for Space Researches* 

**and Its Relationship to Rainfall and** 

**Fire in the Brazilian Tropical Savanna** 

Jorge Alberto Bustamante, Regina Alvalá and Celso von Randow

The Brazilian savanna, named locally Cerrado, is the second largest Brazilian biome, covering approximately two million km2, especially in the Central Highlands (Ratter *et al*., 1997). This biome is composed predominantly of tropical savanna vegetation and is considered as one of the world's biodiversity hotspots, a priority area for biodiversity conservation in the world (Myers et al., 2000). The Cerrado region is considered the last agricultural frontier in the world (Borlaug, 2002), which has been converted in the last 50 years especially for agriculture and pasture purposes, where natural and mainly anthropogenic annual burning is a common practice. Currently, around 50% of natural vegetation in the Cerrado region has been converted to pastures and crops (PROBIO-MMA, 2007).This conversion has impacted the biological diversity, the hydrological cycle, the energy balance, the climate and the carbon dynamics at local and regional scales due to habitat fragmentation, invasive alien species, soil erosion, pollution of aquifers, degradation of ecosystems and changes in fire regimes (Klink & Machado, 2005; Aquino & Miranda, 2008). The knowledge of spatial distribution, temporal dynamics and biophysical characteristics of the vegetation types, are important elements to improve the

understanding of what is the interaction like between vegetation, precipitation and fire.

account the seasonal patterns of the variables involved?

and geographic information systems (GIS) techniques.

**1.1 Seasonality of Cerrado vegetation** 

The objective of this study is to determine the relationship of environmental variables, such as precipitation and fire, with spatial and temporal distribution patterns of main vegetation type of the Brazilian tropical savanna. Thus, we seek to answer the question: how environmental variables, like rain and fire, influence the main vegetation types, like herbaceous, shrubs, deciduous trees and evergreen trees, in the Cerrado biome taking in

In this study, the potential of multi-temporal satellite data, like TRMM data for precipitation, MODIS vegetation indices products for land cover mapping, and others sensors like GOES and MODIS for fire detection is explored by the use of remote sensing

Phenological parameters of vegetation, such as start and end of the growing season, are strongly influenced by atmospheric conditions (like precipitation, temperature and humidity)

**1. Introduction** 

