**Crop Disease and Pest Monitoring by Remote Sensing**

Wenjiang Huang, Juhua Luo, Jingcheng Zhang, Jinling Zhao, Chunjiang Zhao, Jihua Wang, Guijun Yang, Muyi Huang, Linsheng Huang and Shizhou Du *Beijing Research Center for Information Technology in Agriculture, Beijing China* 

### **1. Introduction**

30 Remote Sensing – Applications

Wu, C., Han, X., Niu, Z. & Dong, J. (2010). An evaluation of EO-1 hyperspectral Hyperion

*Environment*, Vol. 85, No. 1, (April 2003), pp. 109-124, ISSN 0034-4257.

data for chlorophyll content and leaf area index estimation. *International Journal of Remote Sensing*, Vol. 31, No. 4, (February 2010), pp. 1079-1086, ISSN 0143-1161 Zarco-Tejada, P.J., Rueda, C.A. & Ustin, S.L. (2003). Water content estimation in vegetation

with MODIS reflectance data and model inversion methods. *Remote Sensing of* 

Plant diseases and pests can affect a wide range of commercial crops, and result in a significant yield loss. It is reported that at least 10% of global food production is lost due to plant diseases (Christou and Twyman, 2004; Strange and Scott, 2005). Excessive pesticides are used for protecting crops from diseases and pests. This not only increases the cost of production, but also raises the danger of toxic residue in agricultural products. Disease and pest control could be more efficient if disease and pest patches within fields can be identified timely and treated locally. This requires obtaining the information of disease infected boundaries in the field as early and accurately as possible. The most common and conventional method is manual field survey. The traditional ground-based survey method requires high labor cost and produces low efficiency. Thus, it is unfeasible for large area. Fortunately, remote sensing technology can provide spatial distribution information of diseases and pests over a large area with relatively low cost. The presence of diseases or insect feedings on plants or canopy surface causes changes in pigment, chemical concentrations, cell structure, nutrient, water uptake, and gas exchange. These changes result in differences in color and temperature of the canopy, and affect canopy reflectance characteristics, which can be detectable by remote sensing (Raikes and Burpee 1998). Therefore, remote sensing provides a harmless, rapid, and cost-effective means of identifying and quantifying crop stress from differences in the spectral characteristics of canopy surfaces affected by biotic and abiotic stress agents.

This chapter introduces some successful studies about detecting and discriminating yellow rust and aphid (economically important disease and pest in winter wheat in China) using field, airborne and satellite remote sensing.

#### **2. Detecting yellow rust of winter wheat by remote sensing**

Yellow rust *(Biotroph Puccinia striiformis),* also known as stripe rust, is a fungal disease of winter wheat *(Triticum aestivum L.).* It produces leaf lesions (pustules), which are yellow in color and tend to be grouped in patches. Yellow rust often occurs in narrow stripes, 2–3 mm wide that run parallel to the leaf veins. Yellow rust is responsible for approximately 73–85%

Crop Disease and Pest Monitoring by Remote Sensing 33

organic matter, total nitrogen, alkali-hydrolysis nitrogen, available phosphorus and available potassium for both growing seasons. Three cultivars of winter wheat used in 2001- 2002 experiment (2002 Exp) were Jingdong8, Jing9428 and Zhongyou9507, while the cultivars used in 2002-2003 (2003 Exp) were Xuezao, 98-100 and Jing411. All the cultivars applied in both growing seasons included erective, middle and loose with respect to the

Growth period Sep 2002-Jun 2003 Sep 2001-Jun 2002

Organic matter 1.42%-1.48% 1.21%-1.32% Total nitrogen 0.08%-0.10% 0.092%-0.124%

phosphorus 20.1-55.4 mg kg-1 25.2-48.3 mg kg-1

Cultivars Xuezao, 98-100, Jing411 Jingdong8, Jing9428,

ha-1 water)

Normal; YR1: 3mg 100-1 ml spores solution; YR2: 9mg 100-1 ml spores solution; YR3: 12mg 100-1 ml spores solution (all treatments applied 200 kg ha-1 nitrogen and 450 m3

Table 1. Basic information of disease inoculation experiment and nutrient stress experiment

For 2002 Exp, six stress treatments of water and nitrogen were applied, and the treatments were based on local conditions, which usually suffered from yellow rust in the northern part

58.6-68.0 mg kg-1 68.8-74.0 mg kg-1

117.6-129.1 mg kg-1 96.6-128.8 mg kg-1

experiment Nutrient stress experiment

Zhongyou9507

450 m3 ha-1 water;

225 m3 ha-1 water;

m3 ha-1 water;

m3 ha-1 water;

m3 ha-1 water;

207, 216, 225, 230, 233 196, 214, 225, 232, 239

ha-1 water;

Normal: 200 kg ha-1 nitrogen,

W-SD: 200 kg ha-1 nitrogen,

W-SED: 200 kg ha-1 nitrogen, 0

N-E: 350 kg ha-1 nitrogen, 450

N-D: 0 kg ha-1 nitrogen, 450 m3

W-SED+N-E: 350 kg ha-1 nitrogen, 0 m3 ha-1 water; W-SED+N-D: 0 kg ha-1 nitrogen, 0

Items Disease inoculation

canopy morphology.

Top soil nutrient status (0-0.3m depth)

Treatments

Spectral reflectance measurements (on day after sowing)

Alkalihydrolysis nitrogen

Available

Rapidly available potassium

of recorded yield losses, and grain quality is also significantly reduced (Li et al. 1989). Consequently, effective monitoring of the incidence and severity of yellow rust in susceptible regions is of great importance to guide the spray of pesticides and to provide data for the local agricultural insurance services. Fortunately, remote sensing technology provides a possible way to detect the incidence and severity of the disease rapidly.

The interaction of electromagnetic radiation with plants varies with the wavelength of the radiation. The same plant leaves may exhibit significant different reflectance depending on the level of health and or vigor (Wooley 1971, West et al. 2003, Luo et al., 2010). Healthy and vigorously growing plant leaves will generally have


The incidence and severity of yellow rust can be monitored according to the differences of spectral characteristics between healthy and disease plants. In this chapter, we will report several successful studies on the detection and identification of yellow rust in winter wheat by remote sensing.

#### **2.1 Detecting and discriminating yellow rust at canopy level**

Hyperspectral remote sensing is one of the advanced and effective techniques in disease monitoring and mapping. However, the difficulty in discriminating a disease from common nutrient stresses largely hampers the practical use of this technique. This is because some common nutrient stresses such as the shortage or overuse of nitrogen or water could have similar variations of biochemical properties and plant morphology, and therefore result in similar spectral responses. However, for the remedial procedures for stressed crops, there is a significant difference between disease and nutrient stresses. For example, applying fungicide to water-stressed crops would lead to a disastrous outcome. Therefore, to discriminate yellow rust from common nutrient stresses is of practical importance to crop growers or landowners.

The specific objectives of this study are to: (1) systematically test the sensitivity and consistency of several commonly used spectral features to yellow rust disease during major growth stages; (2) for those spectral features that are consistently sensitive to yellow rust disease, we will further examine their sensitivity to nutrient stresses to determine whether there are specifically sensitive to yellow rust disease, but insensitive to water and nitrogen stresses.

#### **2.1.1 Materials and methods**

#### **2.1.1.1 Experimental design and field conditions**

The experiments were conducted at Beijing Xiaotangshan Precision Agriculture Experimental Base, in Changping district, Beijing (40º10.6'N, 116º26.3'E) for the growing seasons of 2001-2002 and 2002-2003. Table 1 summarizes the soil properties including

of recorded yield losses, and grain quality is also significantly reduced (Li et al. 1989). Consequently, effective monitoring of the incidence and severity of yellow rust in susceptible regions is of great importance to guide the spray of pesticides and to provide data for the local agricultural insurance services. Fortunately, remote sensing technology

The interaction of electromagnetic radiation with plants varies with the wavelength of the radiation. The same plant leaves may exhibit significant different reflectance depending on the level of health and or vigor (Wooley 1971, West et al. 2003, Luo et al., 2010). Healthy and

1. Low reflectance at visible wavelengths owing to strong absorption by photoactive

2. High reflectance in the near infrared because of multiple scattering at the air-cell

3. Low reflectance in wide wavebands in the short-wave infrared because of absorption by

The incidence and severity of yellow rust can be monitored according to the differences of spectral characteristics between healthy and disease plants. In this chapter, we will report several successful studies on the detection and identification of yellow rust in winter wheat

Hyperspectral remote sensing is one of the advanced and effective techniques in disease monitoring and mapping. However, the difficulty in discriminating a disease from common nutrient stresses largely hampers the practical use of this technique. This is because some common nutrient stresses such as the shortage or overuse of nitrogen or water could have similar variations of biochemical properties and plant morphology, and therefore result in similar spectral responses. However, for the remedial procedures for stressed crops, there is a significant difference between disease and nutrient stresses. For example, applying fungicide to water-stressed crops would lead to a disastrous outcome. Therefore, to discriminate yellow rust from common nutrient stresses is of practical importance to crop

The specific objectives of this study are to: (1) systematically test the sensitivity and consistency of several commonly used spectral features to yellow rust disease during major growth stages; (2) for those spectral features that are consistently sensitive to yellow rust disease, we will further examine their sensitivity to nutrient stresses to determine whether there are specifically sensitive to yellow rust disease, but insensitive to water and nitrogen

The experiments were conducted at Beijing Xiaotangshan Precision Agriculture Experimental Base, in Changping district, Beijing (40º10.6'N, 116º26.3'E) for the growing seasons of 2001-2002 and 2002-2003. Table 1 summarizes the soil properties including

provides a possible way to detect the incidence and severity of the disease rapidly.

vigorously growing plant leaves will generally have

interfaces in the leaf's internal tissue.

by remote sensing.

growers or landowners.

**2.1.1 Materials and methods** 

**2.1.1.1 Experimental design and field conditions** 

stresses.

pigments (chlorophylls, anthocyanins, carotenoids).

**2.1 Detecting and discriminating yellow rust at canopy level** 

water, proteins, and other carbon constituents.

organic matter, total nitrogen, alkali-hydrolysis nitrogen, available phosphorus and available potassium for both growing seasons. Three cultivars of winter wheat used in 2001- 2002 experiment (2002 Exp) were Jingdong8, Jing9428 and Zhongyou9507, while the cultivars used in 2002-2003 (2003 Exp) were Xuezao, 98-100 and Jing411. All the cultivars applied in both growing seasons included erective, middle and loose with respect to the canopy morphology.


Table 1. Basic information of disease inoculation experiment and nutrient stress experiment

For 2002 Exp, six stress treatments of water and nitrogen were applied, and the treatments were based on local conditions, which usually suffered from yellow rust in the northern part

Crop Disease and Pest Monitoring by Remote Sensing 35

Variable Definition Description Literatures

λb Wavelength at Db λ<sup>b</sup> is wavelength position at Db Gong et al., 2002

maximum value of 1st order

λy Wavelength at Dy λ<sup>y</sup> is wavelength position at Dy Gong et al., 2002

λr Wavelength at Dr λr is wavelength position at Dr Gong et al., 2002

In the range of 550nm-750nm Pu et al., 2003;2004 DEP920-1120 In the range of 920nm-1120nm

In the range of 550nm-750nm Pu et al., 2003;2004 WID920-1120 In the range of 920nm-1120nm

In the range of 550nm-750nm

Pu et al., 2003;2004 AREA920-

bands

edge

28 bands

edge

bands

Blue edge covers 490-530nm. Db is a maximum value of 1st order

derivatives within the blue edge of 35

Defined by sum of 1st order derivative values of 35 bands within the blue

Yellow edge covers 550-582nm. Dy is a

derivatives within the yellow edge of

Defined by sum of 1st order derivative values of 28 bands within the yellow

Red edge covers 670-737nm. Dr is a maximum value of 1st order

derivatives within the red edge of 61

Defined by sum of 1st order derivative

values of 61 bands within the red edge Gong et al., 2002

Gong et al., 2002

Gong et al., 2002

Gong et al., 2002

Gong et al., 2002

Gong et al., 2002

**Derivative transformed spectral variables**

Sum of 1st derivative values within blue edge

edge

edge

Sum of 1st derivative values within yellow

Sum of 1st derivative values within red edge

DEP550-750 The depth of the

hull

750 The area of the

wavelength width at half DEP (nm)

absorption feature that is the product of DEP and WID

WID550-750 The full

Maximum value of 1st derivative within red edge

**Continuous removal transformed spectral features**

<sup>1320</sup>In the range of 1070nm-1320nm

<sup>1320</sup>In the range of 1070nm-1320nm

<sup>1120</sup>In the range of 920nm-1120nm

1320 In the range of 1070nm-1320nm

feature minimum relative to the

Maximum value of 1st derivative within blue edge

Maximum value of 1st derivative within yellow

Db

SDb

Dy

SDy

Dr

SDr

DEP1070-

WID1070-

AREA550-

AREA1070-

of China. Each treatment was applied on 0.3 ha area, and the treatments were 200 kg ha-1 nitrogen and 225 m3 ha-1 water (slightly deficient water, W-SD),200 kg ha-1 nitrogen and no irrigation (seriously deficient water, W-SED), 350 kg ha-1 nitrogen and 450 m3 ha-1 water (excessive nitrogen, N-E), no fertilization and 450 m3 ha-1 water (deficient nitrogen, N-D), 350 kg ha-1 nitrogen and no irrigation (seriously deficient water and excessive nitrogen, W-SED+N-E), and no fertilization and no irrigation (seriously deficient water and deficient nitrogen, W-SED+N-D). A 0.3 ha reference area (Normal) was applied with the recommended rate which received 200 kg ha-1 nitrogen and 450 m3 ha-1 water. Three cultivars were evenly distributed in each treatment plot.

For 2003 Exp, according to the National Plant Protection Standard (Li et al. 1989), three levels of concentration of summer spores of yellow rust were applied, and they were 3 mg 100-1 ml-1 (Yellow rust 1, YR1), 9 mg 100-1 ml-1 (Yellow rust 2, YR2) and 12 mg 100-1 ml-1 (Yellow rust 3, YR3), with a dosage of 5 ml spores solution per square meter. The reference area (Normal) that was not inoculated yet was applied with the recommended amount of fungicide to prevent the occasional infection. Each treatment involved 1.2 ha area, with even constitution of three cultivars. All plots in 2003 Exp received the recommended rates of nitrogen (200 kg ha-1) and water (450 m3 ha-1).

#### **2.1.1.2 Canopy spectral measurements**

A high spectral resolution spectrometer, ASD FieldSpec Pro spectrometer (Analytical Spectral Devices, Boulder, CO, USA) fitted with a 25 field of view fore-optic, was used for in-situ measurement of canopy spectral reflectance for both 2002 Exp and 2003 Exp. All canopy spectral measurements were taken from a height of 1.3m above ground (the height of the wheat is 90±3 cm at maturity). Spectra were acquired in the 350-2,500 nm spectral range at a spectral resolution of 3 nm between 350 nm and 1,050 nm, and 10 nm between 1,050 nm and 2,500 nm. A 40 cm × 40 cm BaSO4 calibration panel was used for calculation of reflectance. All irradiance measurements were recorded as an average of 20 scans at an optimized integration time. Prior to subsequent preprocessing, all spectral curves were resampled with 1 nm interval. All measurements were made under clear blue sky conditions between 10:00 and 14:00 (Beijing Local Time).

The spectral measurements were taken 5 times from 196 days after sowing (DAS) to 239 DAS for 2002 Exp, which covered the growth stages of stem elongation, booting, anthesis and milk development. For 2003 Exp, the spectral measurements were taken 5 times from 207 DAS to 233 DAS, which covered the growth stages of booting, anthesis and milk development. The detailed measurement dates for both experiments were given in Table 1. The stem elongation and anthesis stages are essential for the control of yellow rust development, whereas the milk development stage is important for yield loss assessment.

#### **2.1.1.3 Selection of spectral features**

The spectral features that we adopted were related to several commonly used vegetation indices (VIs), which were proved to be sensitive to variations of pigments and stresses. Furthermore, in order to conduct a thorough investigation of various types of spectral features, we also included a number of spectral features that were based on derivative transformation and continuum removal transformation (Gong et al. 2002; Pu et al. 2003;2004). Therefore, the total 38 spectral features are shown in Table 2.

of China. Each treatment was applied on 0.3 ha area, and the treatments were 200 kg ha-1 nitrogen and 225 m3 ha-1 water (slightly deficient water, W-SD),200 kg ha-1 nitrogen and no irrigation (seriously deficient water, W-SED), 350 kg ha-1 nitrogen and 450 m3 ha-1 water (excessive nitrogen, N-E), no fertilization and 450 m3 ha-1 water (deficient nitrogen, N-D), 350 kg ha-1 nitrogen and no irrigation (seriously deficient water and excessive nitrogen, W-SED+N-E), and no fertilization and no irrigation (seriously deficient water and deficient nitrogen, W-SED+N-D). A 0.3 ha reference area (Normal) was applied with the recommended rate which received 200 kg ha-1 nitrogen and 450 m3 ha-1 water. Three

For 2003 Exp, according to the National Plant Protection Standard (Li et al. 1989), three levels of concentration of summer spores of yellow rust were applied, and they were 3 mg 100-1 ml-1 (Yellow rust 1, YR1), 9 mg 100-1 ml-1 (Yellow rust 2, YR2) and 12 mg 100-1 ml-1 (Yellow rust 3, YR3), with a dosage of 5 ml spores solution per square meter. The reference area (Normal) that was not inoculated yet was applied with the recommended amount of fungicide to prevent the occasional infection. Each treatment involved 1.2 ha area, with even constitution of three cultivars. All plots in 2003 Exp received the recommended rates of

A high spectral resolution spectrometer, ASD FieldSpec Pro spectrometer (Analytical Spectral Devices, Boulder, CO, USA) fitted with a 25 field of view fore-optic, was used for in-situ measurement of canopy spectral reflectance for both 2002 Exp and 2003 Exp. All canopy spectral measurements were taken from a height of 1.3m above ground (the height of the wheat is 90±3 cm at maturity). Spectra were acquired in the 350-2,500 nm spectral range at a spectral resolution of 3 nm between 350 nm and 1,050 nm, and 10 nm between 1,050 nm and 2,500 nm. A 40 cm × 40 cm BaSO4 calibration panel was used for calculation of reflectance. All irradiance measurements were recorded as an average of 20 scans at an optimized integration time. Prior to subsequent preprocessing, all spectral curves were resampled with 1 nm interval. All measurements were made under clear blue sky conditions

The spectral measurements were taken 5 times from 196 days after sowing (DAS) to 239 DAS for 2002 Exp, which covered the growth stages of stem elongation, booting, anthesis and milk development. For 2003 Exp, the spectral measurements were taken 5 times from 207 DAS to 233 DAS, which covered the growth stages of booting, anthesis and milk development. The detailed measurement dates for both experiments were given in Table 1. The stem elongation and anthesis stages are essential for the control of yellow rust development, whereas the milk development stage is important for yield loss assessment.

The spectral features that we adopted were related to several commonly used vegetation indices (VIs), which were proved to be sensitive to variations of pigments and stresses. Furthermore, in order to conduct a thorough investigation of various types of spectral features, we also included a number of spectral features that were based on derivative transformation and continuum removal transformation (Gong et al. 2002; Pu et al.

2003;2004). Therefore, the total 38 spectral features are shown in Table 2.

cultivars were evenly distributed in each treatment plot.

nitrogen (200 kg ha-1) and water (450 m3 ha-1).

between 10:00 and 14:00 (Beijing Local Time).

**2.1.1.3 Selection of spectral features** 

**2.1.1.2 Canopy spectral measurements** 


Crop Disease and Pest Monitoring by Remote Sensing 37

Variable Definition Description Literatures

[(R712+R752)/2]-R732

**As the first step, a**ll spectra were processed with the following transformation to suppress possible difference in illumination. The spectral regions with wavelength of 1330-1450 nm, 1770-2000 nm and 2400-2500 nm were removed due to strong absorption by water vapor. We then normalized the spectral curves by dividing the mean band reflectance of the curve

> 1 <sup>1</sup> ( )

*i*

where *Refi'* is the normalized reflectance for bandi; *Refi* is the original reflectance of the band; n is the total number of bands. Fig. 1(a) shows a plot of unnormalized *Refi* versus band wavelength for six observations (three YR3 curves and three Normal curves) on 233 DAS. Fig. 1(b) shows the corresponding curves in Fig.1(a) after normalization. The normalization clearly separated the diseased spectra from the normal spectra especially over the near infrared region (approximately from 770 nm to 1300 nm). The benefit of eliminating spectral difference caused

As shown in Table 1, although both experiments conducted in five growth stages in 2002 and 2003, most measurement dates were not consistent, except for 255 DAS. Hence, to improve the comparability of two datasets, we adapted the 2002 Exp data to match the dates

*i*

*Ref* 

*i*

n

*i n*

*Ref Ref*

by the change of illumination conditions was also mentioned by Yu et al. (1999).

**Normalization of the difference in measuring dates** 

Stress Index (R802+R547)/(R1657+R682) Galvão et al., 2005

(R860-R1640)/(R860+R1640) Fensholt and

(R701-R671)-0.2(R701-R549)]/(R701/R671) Daughry et al.,

Hunt and rock, 1989; Ceccato et

Sandholt, 2003

Peñuelas et al.,

Merton and Huntington, 1999

2000

1997

al., 2001

DSWI Disease Water

MSI Moisture Stress

Shortwave Infrared Water Stress Index

Red-Edge Vegetation Stress

Modified Chlorophyll Absorption in Reflectance Index

WI Water Index R900/R970

**Aggregating spectral reflectance data** 

Table 2. Definitions of spectral features used in this study

**2.1.1.4 Preprocessing and normalization of spectral reflectance data** 

(Yu et al., 1999). The normalized reflectance for the bandi is given as:

Index

SIWSI

RVSI

MCARI

Index R1600/R819


Variable Definition Description Literatures

(RNIR-RR)/(RNIR+RR), where RNIR indicates 775-825nm, RR indicates 650nm-700nm, that include most key

reflectance index (R570-R670)/(R570+R670) Filella et al., 1995

3\*[( R700- R670)-0.2\*( R700- R550)\*( R700/

reflectance index (R550-R531)/(R550+R531) Gamon et al., 1992

Reflectance Index ARI=(R550)-1-(R700)-1 Gitelson et al.,

(|(a670+R670+b)|/(a2+1)1/2)x(R700/R670)

Vegetation Index 0.5[120(R750-R550)-200(R670-R550)]

(R850-R680)/(R850+R680) Thenkabail et al.,

(R531-R570)/(R531+R570) Gamon et al., 1992

(R800-R445)/(R800-R680) Peñuelas et al.,

(R680-R430)/(R680+R430) Peñuelas et al.,

a = (R700-R550)/150, b = R550-(a x 550) Kim et al., 1994

Ratio (R800/R670-1)/(R800/R670+1)1/2

pigments

R670)]

Reflectance Index (R680-R500)/R750

Zarco-Tejada et al.,

Rouse et al., 1973

Haboudane et al.,

Merzlyak et al.,

2005

2004

2000

2002

1995

1999

1994

2001

Broge and Leblanc, 2000; Haboudane et al., 2004

Chen, 1996; Haboudane et al.,

**VI-based variables**

NDVI

NBNDVI

PRI

TCARI

SIPI

NPCI

CARI

NRI Nitrogen

GI Greenness Index R554/R677

Normalized Difference Vegetation Index

Narrow-band normalised difference vegetation index

Photochemical Physiological Reflectance Index

The transformed chlorophyll Absorption and Reflectance Index

Structural Independent Pigment Index

PSRI Plant Senescence

PhRI The Physiological

Index

ARI Anthocyanin

TVI Triangular

Normalized Pigment

Chlorophyll Absorption Ratio

Index

Chlorophyll ratio

MSR Modified Simple


Table 2. Definitions of spectral features used in this study

#### **2.1.1.4 Preprocessing and normalization of spectral reflectance data**

#### **Aggregating spectral reflectance data**

**As the first step, a**ll spectra were processed with the following transformation to suppress possible difference in illumination. The spectral regions with wavelength of 1330-1450 nm, 1770-2000 nm and 2400-2500 nm were removed due to strong absorption by water vapor. We then normalized the spectral curves by dividing the mean band reflectance of the curve (Yu et al., 1999). The normalized reflectance for the bandi is given as:

$$Ref'\_i = \frac{Ref\_i}{\frac{1}{\mathbf{n}} \left(\sum\_{i=1}^n Ref\_i\right)}$$

where *Refi'* is the normalized reflectance for bandi; *Refi* is the original reflectance of the band; n is the total number of bands. Fig. 1(a) shows a plot of unnormalized *Refi* versus band wavelength for six observations (three YR3 curves and three Normal curves) on 233 DAS. Fig. 1(b) shows the corresponding curves in Fig.1(a) after normalization. The normalization clearly separated the diseased spectra from the normal spectra especially over the near infrared region (approximately from 770 nm to 1300 nm). The benefit of eliminating spectral difference caused by the change of illumination conditions was also mentioned by Yu et al. (1999).

#### **Normalization of the difference in measuring dates**

As shown in Table 1, although both experiments conducted in five growth stages in 2002 and 2003, most measurement dates were not consistent, except for 255 DAS. Hence, to improve the comparability of two datasets, we adapted the 2002 Exp data to match the dates

Crop Disease and Pest Monitoring by Remote Sensing 39

of 2003 Exp, by using a linear interpolation method. The reflectance curve of a certain date could be obtained based on the spectra from the adjacent data before and after the measurement date (using days after sowing as a time scale). Each band of the spectra should

> *current before before after after before*

where *Refcurrent* represents the reflectance transformed from the date corresponding to an ideal date in 2003 Exp; *Refbefore* and *Refafter* represent reflectances, respectively, from *DASbefore* and *DASafter*; *DAScurrent* indicates an ideal date in 2003 Exp while *DASbefore* and *DASafter* are the

Fig. 2 provides an example of the progress of the normalization of measurement dates. The averaged reflectance at central wavelengths of green band (560 nm) and near-infrared band (860 nm) of Landsat-5 TM for normal samples were plotted against the measured dates in both 2002 Exp and 2003 Exp. The date normalized reflectance values were marked as triangle symbol in the graph. Through this step, the datasets collected in these two years could be considered as acquired in the same dates, which thereby facilitated the subsequent

Adaptation of average reflectance of normal samples at 560 nm (central wavelengths of green band of Landsat-5 TM) and 860 nm (central wavelengths of near-infrared band of Landsat-5 TM) to match the

dates of 2003 Exp, by using a linear interpolation method

Fig. 2. An example for normalization of measuring dates

adjacent dates in 2002 Exp before and after the ideal date in 2003 Exp.

*DAS DAS Ref Ref Ref Ref DAS DAS* 

( ) *current before*

be processed as:

comparisons and analysis.

(b) Normalized spectra on 233 days after sowing

Fig. 1. Comparison between original spectra and normalized ones

(a) Original spectra on 233 days after sowing

(b) Normalized spectra on 233 days after sowing

Fig. 1. Comparison between original spectra and normalized ones

of 2003 Exp, by using a linear interpolation method. The reflectance curve of a certain date could be obtained based on the spectra from the adjacent data before and after the measurement date (using days after sowing as a time scale). Each band of the spectra should be processed as:

$$Ref\_{current} = Ref\_{before} - \frac{DAS\_{current} - DAS\_{before}}{DAS\_{after} - DAS\_{before}} (Ref\_{before} - Ref\_{after})$$

where *Refcurrent* represents the reflectance transformed from the date corresponding to an ideal date in 2003 Exp; *Refbefore* and *Refafter* represent reflectances, respectively, from *DASbefore* and *DASafter*; *DAScurrent* indicates an ideal date in 2003 Exp while *DASbefore* and *DASafter* are the adjacent dates in 2002 Exp before and after the ideal date in 2003 Exp.

Fig. 2 provides an example of the progress of the normalization of measurement dates. The averaged reflectance at central wavelengths of green band (560 nm) and near-infrared band (860 nm) of Landsat-5 TM for normal samples were plotted against the measured dates in both 2002 Exp and 2003 Exp. The date normalized reflectance values were marked as triangle symbol in the graph. Through this step, the datasets collected in these two years could be considered as acquired in the same dates, which thereby facilitated the subsequent comparisons and analysis.

Adaptation of average reflectance of normal samples at 560 nm (central wavelengths of green band of Landsat-5 TM) and 860 nm (central wavelengths of near-infrared band of Landsat-5 TM) to match the dates of 2003 Exp, by using a linear interpolation method

Fig. 2. An example for normalization of measuring dates

Crop Disease and Pest Monitoring by Remote Sensing 41

Fig. 3. Ratios of spectra for normalization with different years and varieties

**2.1.2.3 One way ANOVA of four disease sensitive spectral features** 

rational comparisons among different treatments.

**2.1.2.2 Spectral responses to different forms of stresses** 

from variation of illumination and different measurement dates, etc., and enabled more

The result of ANOVA between normal samples and different forms of stress samples indicated that all spectral features had a response (defined as *p*-value<0.05) to at least one type of stresses at one growth stage, except for the WID1070-1320, which had no response to any form of stresses at all growth stages. Total 37 spectral features responded to water associated stresses (W-SD, W-SED, W-SED+N-E, W-SED+N-D) at least at one growth stage, followed by 35 spectral features to yellow rust disease, whereas only15 spectral features had a response to solely nitrogen stress (N-E, N-D). As summarized in Table 3, most spectral features were sensitive to yellow rust infection at least at one growth stage, except for λb, λ<sup>r</sup> and WID1070-1320. In addition, most spectral features tended to be more sensitive at later growth stages than at the early stages. For example, several features such as DEP920-1120, AREA920-1120, Dy, GI, NDVI and Triangular Vegetation Index (TVI) only had a response to yellow rust at the last growth stage in our study (233 DAS). However, for the sake of diagnosis, the spectral features with a consistent response to yellow rust during the important growing period would be much more valuable. Therefore, those spectral features that were sensitive to the yellow rust at 4 out of 5 growth stages were selected as candidates for disease diagnosis. This yielded four vegetation indices (VIs): PRI, PhRI, NPCI and ARI.

Particularly for the four identified VIs that closely associated with yellow rust disease, a throughout one way ANOVA was conducted to compare their differences between the

#### **Normalization of the difference from cultivars and soil backgrounds**

The canopy spectra of winter wheat were not only supposed to respond to stresses, but are also determined and influenced by several other aspects such as cultivars and soil properties. Although the both 2002 Exp and 2003 Exp were conducted in the same fields that had approximately identical climate and environmental conditions, the difference in cultivars and soil properties between 2002 Exp and 2003 Exp should not be ignored (Table 1). To minimize this discrepancy, we calculated a ratio spectral curve for each of measured dates (after the normalization of the measuring dates) by the averaged spectral curve from normal samples in 2002 Exp divided by the averaged spectral curve from normal samples in 2003 Exp, resulting in a total of five ratio curves corresponding to each growth stage (Fig. 3). After that, all the spectral data measured at different growth stages were multiplied by the corresponding ratio curves to yield a set of normalized spectra. It should be pointed out that the present normalization processing to raw spectral measurements will only enhance the comparability between the 2002 Exp and 2003 Exp with little change in internal relations among different treatments because all the spectral data at one growth stage were processed with the same ratio curve. The ultimate goal of all these preprocessing and normalization steps above is to mitigate effects of the variation of illumination conditions, measurement dates, cultivars and soil properties between the 2002 Exp and 2003 Exp on target spectra.

#### **2.1.1.5 Spectral features calculation and statistical analysis**

With the spectra normalized using the methods above, we calculated 38 spectral features. An analysis of variance (ANOVA) was employed to investigate the spectral differences between the normal samples and all forms of stressed samples. Firstly, on different measured dates, both the yellow rust disease data and nutrient stressed data were compared with the normal data by ANOVA. For those spectral features that were consistently sensitive to yellow rust disease, we not only tested their differences between the normal treatment and different forms of stresses, but also tested the differences between various kinds of nutrient stresses and varying levels of disease stresses with ANOVA. Statistical analyses were conducted using SPSS 13.0 procedure.

#### **2.1.2 Results**

#### **2.1.2.1 Spectra after normalizations**

The spectral ratio curves in Fig 3 reflect the deviations between 2002 Exp and 2003 Exp's reflectance datasets at different wavelength positions. The ratio value close to 1.0 indicates no difference in reflectance exists between the two years. Generally, the ratio values ranged from 0.7 to 1.3, with an uneven distribution along the wavelength axis (Fig 3). The ratio tended to deviate from 1.0 in the regions of 350 - 730 nm, 1450 - 1570 and 2000 - 2400 nm, but stayed around 1.0 in the regions of 730 - 1330 nm and 1570 - 1770 nm. To assess the improvement in comparability, we examined the difference of normalized datasets of normal samples between 2002 Exp and 2003 Exp through an ANOVA with all 38 spectral features. The result showed that the differences of all spectral features were insignificant at all growth stages (*p*-value>0.05), with an average *p*-value (for all measuring dates) of 0.94, indicating a relatively high level of similarity between two datasets. Therefore, we confirmed that such normalization processes minimized the spectral difference originated

The canopy spectra of winter wheat were not only supposed to respond to stresses, but are also determined and influenced by several other aspects such as cultivars and soil properties. Although the both 2002 Exp and 2003 Exp were conducted in the same fields that had approximately identical climate and environmental conditions, the difference in cultivars and soil properties between 2002 Exp and 2003 Exp should not be ignored (Table 1). To minimize this discrepancy, we calculated a ratio spectral curve for each of measured dates (after the normalization of the measuring dates) by the averaged spectral curve from normal samples in 2002 Exp divided by the averaged spectral curve from normal samples in 2003 Exp, resulting in a total of five ratio curves corresponding to each growth stage (Fig. 3). After that, all the spectral data measured at different growth stages were multiplied by the corresponding ratio curves to yield a set of normalized spectra. It should be pointed out that the present normalization processing to raw spectral measurements will only enhance the comparability between the 2002 Exp and 2003 Exp with little change in internal relations among different treatments because all the spectral data at one growth stage were processed with the same ratio curve. The ultimate goal of all these preprocessing and normalization steps above is to mitigate effects of the variation of illumination conditions, measurement dates, cultivars and soil properties

With the spectra normalized using the methods above, we calculated 38 spectral features. An analysis of variance (ANOVA) was employed to investigate the spectral differences between the normal samples and all forms of stressed samples. Firstly, on different measured dates, both the yellow rust disease data and nutrient stressed data were compared with the normal data by ANOVA. For those spectral features that were consistently sensitive to yellow rust disease, we not only tested their differences between the normal treatment and different forms of stresses, but also tested the differences between various kinds of nutrient stresses and varying levels of disease stresses with ANOVA. Statistical

The spectral ratio curves in Fig 3 reflect the deviations between 2002 Exp and 2003 Exp's reflectance datasets at different wavelength positions. The ratio value close to 1.0 indicates no difference in reflectance exists between the two years. Generally, the ratio values ranged from 0.7 to 1.3, with an uneven distribution along the wavelength axis (Fig 3). The ratio tended to deviate from 1.0 in the regions of 350 - 730 nm, 1450 - 1570 and 2000 - 2400 nm, but stayed around 1.0 in the regions of 730 - 1330 nm and 1570 - 1770 nm. To assess the improvement in comparability, we examined the difference of normalized datasets of normal samples between 2002 Exp and 2003 Exp through an ANOVA with all 38 spectral features. The result showed that the differences of all spectral features were insignificant at all growth stages (*p*-value>0.05), with an average *p*-value (for all measuring dates) of 0.94, indicating a relatively high level of similarity between two datasets. Therefore, we confirmed that such normalization processes minimized the spectral difference originated

**Normalization of the difference from cultivars and soil backgrounds** 

between the 2002 Exp and 2003 Exp on target spectra.

analyses were conducted using SPSS 13.0 procedure.

**2.1.2.1 Spectra after normalizations** 

**2.1.2 Results** 

**2.1.1.5 Spectral features calculation and statistical analysis** 

Fig. 3. Ratios of spectra for normalization with different years and varieties

from variation of illumination and different measurement dates, etc., and enabled more rational comparisons among different treatments.

#### **2.1.2.2 Spectral responses to different forms of stresses**

The result of ANOVA between normal samples and different forms of stress samples indicated that all spectral features had a response (defined as *p*-value<0.05) to at least one type of stresses at one growth stage, except for the WID1070-1320, which had no response to any form of stresses at all growth stages. Total 37 spectral features responded to water associated stresses (W-SD, W-SED, W-SED+N-E, W-SED+N-D) at least at one growth stage, followed by 35 spectral features to yellow rust disease, whereas only15 spectral features had a response to solely nitrogen stress (N-E, N-D). As summarized in Table 3, most spectral features were sensitive to yellow rust infection at least at one growth stage, except for λb, λ<sup>r</sup> and WID1070-1320. In addition, most spectral features tended to be more sensitive at later growth stages than at the early stages. For example, several features such as DEP920-1120, AREA920-1120, Dy, GI, NDVI and Triangular Vegetation Index (TVI) only had a response to yellow rust at the last growth stage in our study (233 DAS). However, for the sake of diagnosis, the spectral features with a consistent response to yellow rust during the important growing period would be much more valuable. Therefore, those spectral features that were sensitive to the yellow rust at 4 out of 5 growth stages were selected as candidates for disease diagnosis. This yielded four vegetation indices (VIs): PRI, PhRI, NPCI and ARI.

#### **2.1.2.3 One way ANOVA of four disease sensitive spectral features**

Particularly for the four identified VIs that closely associated with yellow rust disease, a throughout one way ANOVA was conducted to compare their differences between the

Crop Disease and Pest Monitoring by Remote Sensing 43

normal sample and various kinds of stressed samples. Moreover, their differences among each pairs of stress forms were also compared. We conducted this ANOVA based on the data on 207 DAS, 225 DAS and 233 DAS respectively, which were essential growth stages for carrying out fungicide spraying and yield loss assessing procedures. In addition to the *p*value of ANOVA, we also provided the change direction of spectral features. Positive sign indicates the average spectral feature value of diseased or nutrient stressed samples is greater than that of normal samples, and negative sign indicates the opposite cases to the positive sign. As shown in Table 4, it was observed that for the treatments of N-E and N-D, all four VIs failed to show any response at all growth stages. For the results of other treatments, the responses of four VIs behaved in a varied pattern at three growth stages.

For the results on 207 DAS (Table 4a), compared to the normal samples, the NPCI and ARI had responses to all three levels of yellow rust treatments (YR 1, YR 2, YR 3), and appeared to be more sensitive than PRI and PhRI. For nutrient stresses, the PRI, NPCI and ARI were sensitive to W-SED and W-SED+N-E treatments. Among them, NPCI and ARI showed stronger responses (*p*-value<0.01) to W-SD, W-SED, W-SED+N-E and W-SED+N-D treatments than the other two VIs. For the comparisons between diseased samples and nutrient stressed samples, significant differences between W-SED and W-SED+N-E treatments and YR2 and YR3 treatments were identified for PRI, NPCI and ARI. Moreover, the change directions of the three VIs for diseased and nutrient stressed samples were identical. At this 207 DAS growth stage, PhRI did not show a significant response to any of three levels of disease treatments, but responded to W-SD, W-SED and W-SED+N-E treatments. It is interesting that the change direction of diseased samples of PhRI was contrary to that of the nutrient stressed samples, suggesting a discriminating potential of the

For the results on 225 DAS (Table 4b), compared to the normal samples, all four VIs revealed a clear response to level 2 and level 3 of yellow rust treatments (YR2, YR3). For nutrient stresses, PRI, NPCI and ARI also appeared to be sensitive to W-SD, W-SED, W-SED+N-E and W-SED+N-D treatments. However, PhRI was insensitive to all nutrient stresses. In addition, when we looked at the difference of those VIs between diseased samples and nutrient stressed samples, only PhRI showed clear differences between YR2 and YR3 treatments and W-SD, W-SED, W-SED+N-E, and W-SED+N-D treatments. Although a significant difference between YR3 treatment and W-SED treatment also existed for ARI and NPCI, the change directions of both treatments were identical. However, for PhRI, the change directions of all levels of disease treatments were different from those of the nutrient

For the results on 233 DAS (Table 4c), with further development of disease symptoms, compared to the normal samples, all four indices showed responses to all three levels of disease treatments. Comparing to YR1 treatment, the four VIs had shown a stronger significant level (*p*-value<0.01) for YR2, YR3 treatments. For nutrient stresses, PRI, NPCI and ARI exhibited clear responses to W-SED, W-SED+N-E and W-SED+N-D treatments as well. For comparisons between diseased and nutrient stressed samples, PRI and NPCI appeared to be significantly different between YR2 and YR3 treatments and W-SD treatment. However, the change directions of both treatments were identical. Unlike the other three VIs, PhRI remained insensitive to the nutrient stresses, but was significantly different among all levels of disease treatments (YR1, YR2, and YR3) and all forms of nutrient stresses. More

index.

stress treatments.


Table 3. Responses of spectral features to yellow rust

DEP550-770 √ √ √ AREA550-770 √ √ √ WID550-770 √ √ √ DEP920-1120 √ AREA920-1120 √ WID920-1120 √ DEP1070-1320 √ AREA1070-1320 √

SDb √ √ √ Dy √ λy √ SDy √

SDr √ √ GI √ MSR √ √ NDVI √ NBNDVI √ √ NRI √ PRI √ √ √ √

SIPI √ PSRI √ √ √ PhRI √ √ √ √ NPCI √ √ √ √ ARI √ √ √ √ TVI √ CARI √ √ √ DSWI √ MSI √ SIWSI √

207 216 225 230 233

Spectral features Days after sowing

Db √ √

Dr √

TCARI √ √

RVSI √ √

Table 3. Responses of spectral features to yellow rust

MCARI √ √ √ WI √ normal sample and various kinds of stressed samples. Moreover, their differences among each pairs of stress forms were also compared. We conducted this ANOVA based on the data on 207 DAS, 225 DAS and 233 DAS respectively, which were essential growth stages for carrying out fungicide spraying and yield loss assessing procedures. In addition to the *p*value of ANOVA, we also provided the change direction of spectral features. Positive sign indicates the average spectral feature value of diseased or nutrient stressed samples is greater than that of normal samples, and negative sign indicates the opposite cases to the positive sign. As shown in Table 4, it was observed that for the treatments of N-E and N-D, all four VIs failed to show any response at all growth stages. For the results of other treatments, the responses of four VIs behaved in a varied pattern at three growth stages.

For the results on 207 DAS (Table 4a), compared to the normal samples, the NPCI and ARI had responses to all three levels of yellow rust treatments (YR 1, YR 2, YR 3), and appeared to be more sensitive than PRI and PhRI. For nutrient stresses, the PRI, NPCI and ARI were sensitive to W-SED and W-SED+N-E treatments. Among them, NPCI and ARI showed stronger responses (*p*-value<0.01) to W-SD, W-SED, W-SED+N-E and W-SED+N-D treatments than the other two VIs. For the comparisons between diseased samples and nutrient stressed samples, significant differences between W-SED and W-SED+N-E treatments and YR2 and YR3 treatments were identified for PRI, NPCI and ARI. Moreover, the change directions of the three VIs for diseased and nutrient stressed samples were identical. At this 207 DAS growth stage, PhRI did not show a significant response to any of three levels of disease treatments, but responded to W-SD, W-SED and W-SED+N-E treatments. It is interesting that the change direction of diseased samples of PhRI was contrary to that of the nutrient stressed samples, suggesting a discriminating potential of the index.

For the results on 225 DAS (Table 4b), compared to the normal samples, all four VIs revealed a clear response to level 2 and level 3 of yellow rust treatments (YR2, YR3). For nutrient stresses, PRI, NPCI and ARI also appeared to be sensitive to W-SD, W-SED, W-SED+N-E and W-SED+N-D treatments. However, PhRI was insensitive to all nutrient stresses. In addition, when we looked at the difference of those VIs between diseased samples and nutrient stressed samples, only PhRI showed clear differences between YR2 and YR3 treatments and W-SD, W-SED, W-SED+N-E, and W-SED+N-D treatments. Although a significant difference between YR3 treatment and W-SED treatment also existed for ARI and NPCI, the change directions of both treatments were identical. However, for PhRI, the change directions of all levels of disease treatments were different from those of the nutrient stress treatments.

For the results on 233 DAS (Table 4c), with further development of disease symptoms, compared to the normal samples, all four indices showed responses to all three levels of disease treatments. Comparing to YR1 treatment, the four VIs had shown a stronger significant level (*p*-value<0.01) for YR2, YR3 treatments. For nutrient stresses, PRI, NPCI and ARI exhibited clear responses to W-SED, W-SED+N-E and W-SED+N-D treatments as well. For comparisons between diseased and nutrient stressed samples, PRI and NPCI appeared to be significantly different between YR2 and YR3 treatments and W-SD treatment. However, the change directions of both treatments were identical. Unlike the other three VIs, PhRI remained insensitive to the nutrient stresses, but was significantly different among all levels of disease treatments (YR1, YR2, and YR3) and all forms of nutrient stresses. More

Crop Disease and Pest Monitoring by Remote Sensing 45

Combining with a dataset of yellow rust disease inoculation and a dataset of various forms of nutrient stress treatments, we examined the responses of 38 commonly used spectral features at five important growth stages from booting stage to milk development stage using a one-way analysis of variance (ANOVA). There were 37 spectral features sensitive to water associated stresses, 35 spectral features sensitive to yellow rust disease and only 15 spectral features sensitive to sole nitrogen stresses in at least one growth stage. It was observed that more spectral features appeared to have a response to yellow rust disease at later growth stages. A throughout ANOVA was conducted particularly on PRI, PhRI, NPCI and ARI, which showed a consistent response to yellow rust disease at 4 out of 5 growth stages. However, PRI, NPCI and ARI were also responsible for water associated stresses, suggesting a risk of confusion in detecting yellow rust disease. Only PhRI was sensitive to yellow rust disease, but insensitive to different forms of nutrient stresses. The discriminative response of PhRI could provide a means of identifying and detecting yellow rust disease under complicated farmland circumstances. This finding can serve the basis of remote

**2.1.3 Conclusion** 

sensing system for detecting yellow rust disease.

**2.2.1.1 Experimental design and field conditions** 

disease index (DI) was then calculated using (Li et al. 1989):

using hyperspectral imagery.

**2.2.1 Materials and methods** 

validate the models developed.

**2.2.1.2 Inspection of disease severity** 

were randomly selected for check.

**2.2.1.3 Canopy spectral measurements** 

**2.2 Detecting yellow rust using field and airborne hyperspectral data** 

The aim of this study was to evaluate the accuracy of the spectro-optical, photochemical reflectance index (PRI) for quantifying the disease index (DI) of yellow rust in wheat using in-situ spectral reflectance measurements, and its applicability in the detection of the disease

Experimental design and field conditions was same as 1.1.1. Experimental data from 2002 Exp were used to establish the statistical models, and the data for 2003 Exp were used to

To quantify the severity of the disease of yellow rust, the leaves of plants were grouped into one of 9 classifications of disease incidence (x): 0,1, 10, 20, 30, 45, 60, 80 and 100% covered by rust. 0% represented no incidence of yellow rust, and 100% was the greatest incidence. The

> (%) <sup>100</sup> *x f DI n f*

where *f* is the total number of leaves of each degree of disease severity and *n* is the degree of disease severity observed (in this work, n ranged from 0 to 8). In each plot, 20 individuals

The method of canopy spectral measurements and data was same as the part 1.1.1.2 above.


#### (a) 207 DAS


#### (b) 225 DAS


#### (c) 233 DAS

\*mean difference is significant at 0.950 confidence level; \*\*mean difference is significant at 0.990 confidence level;\*\*\* mean difference is significant at 0.999 confidence level. (+) means the average spectral feature value of diseased or nutrient stressed samples greater than that of normal samples; or means the average spectral feature value of nutrient stressed samples greater than that of diseased samples; (-) means the opposite cases to the case of (+). The definitions of treatments are as follows: "Normal" represents normal samples; "W-SD" represents samples treated with slightly deficient water; "W-SED" represents samples treated with seriously deficient water; "N-E" represents samples treated with excessive nitrogen; "N-D" represents samples treated with deficient nitrogen; "W-SED+N-E" represents samples treated with seriously deficient water and excessive nitrogen; "W-SED+N-D" represents samples treated with seriously deficient water and deficient nitrogen

Table 4. ANOVA for four VIs separately on 207 DAS, 225 DAS and 233 DAS

importantly for the PhRI, the change directions of diseased samples were opposite to those of nutrient stressed samples throughout the entire analysis.

In summary, all four VIs showed a significant sensitivity to yellow rust disease on 207 DAS, 225 DAS and 233 DAS. However, most of them also appeared to be sensitive to water associated stresses to a varing extent, except for PhRI, which was only sensitive to disease yet insensitive to any forms of nutrient stresses on 225 DAS and 233 DAS. More importantly, the change directions of PhRI to disease treatments were always opposite to those to the nutrient stress treatments at all relevant growth stages. This further confirmed the discriminating characteristic of PhRI.

#### **2.1.3 Conclusion**

44 Remote Sensing – Applications

\*mean difference is significant at 0.950 confidence level; \*\*mean difference is significant at 0.990 confidence level;\*\*\* mean difference is significant at 0.999 confidence level. (+) means the average spectral feature value of diseased or nutrient stressed samples greater than that of normal samples; or means the average spectral feature value of nutrient stressed samples greater than that of diseased samples; (-) means the opposite cases to the case of (+). The definitions of treatments are as follows: "Normal" represents normal samples; "W-SD" represents samples treated with slightly deficient water; "W-SED" represents samples treated with seriously deficient water; "N-E" represents samples treated with excessive nitrogen; "N-D" represents samples treated with deficient nitrogen; "W-SED+N-E" represents samples treated with seriously deficient water and excessive nitrogen; "W-SED+N-D"

importantly for the PhRI, the change directions of diseased samples were opposite to those

In summary, all four VIs showed a significant sensitivity to yellow rust disease on 207 DAS, 225 DAS and 233 DAS. However, most of them also appeared to be sensitive to water associated stresses to a varing extent, except for PhRI, which was only sensitive to disease yet insensitive to any forms of nutrient stresses on 225 DAS and 233 DAS. More importantly, the change directions of PhRI to disease treatments were always opposite to those to the nutrient stress treatments at all relevant growth stages. This further confirmed the

represents samples treated with seriously deficient water and deficient nitrogen Table 4. ANOVA for four VIs separately on 207 DAS, 225 DAS and 233 DAS

of nutrient stressed samples throughout the entire analysis.

discriminating characteristic of PhRI.

(a) 207 DAS

(b) 225 DAS

(c) 233 DAS

Combining with a dataset of yellow rust disease inoculation and a dataset of various forms of nutrient stress treatments, we examined the responses of 38 commonly used spectral features at five important growth stages from booting stage to milk development stage using a one-way analysis of variance (ANOVA). There were 37 spectral features sensitive to water associated stresses, 35 spectral features sensitive to yellow rust disease and only 15 spectral features sensitive to sole nitrogen stresses in at least one growth stage. It was observed that more spectral features appeared to have a response to yellow rust disease at later growth stages. A throughout ANOVA was conducted particularly on PRI, PhRI, NPCI and ARI, which showed a consistent response to yellow rust disease at 4 out of 5 growth stages. However, PRI, NPCI and ARI were also responsible for water associated stresses, suggesting a risk of confusion in detecting yellow rust disease. Only PhRI was sensitive to yellow rust disease, but insensitive to different forms of nutrient stresses. The discriminative response of PhRI could provide a means of identifying and detecting yellow rust disease under complicated farmland circumstances. This finding can serve the basis of remote sensing system for detecting yellow rust disease.

#### **2.2 Detecting yellow rust using field and airborne hyperspectral data**

The aim of this study was to evaluate the accuracy of the spectro-optical, photochemical reflectance index (PRI) for quantifying the disease index (DI) of yellow rust in wheat using in-situ spectral reflectance measurements, and its applicability in the detection of the disease using hyperspectral imagery.

#### **2.2.1 Materials and methods**

#### **2.2.1.1 Experimental design and field conditions**

Experimental design and field conditions was same as 1.1.1. Experimental data from 2002 Exp were used to establish the statistical models, and the data for 2003 Exp were used to validate the models developed.

#### **2.2.1.2 Inspection of disease severity**

To quantify the severity of the disease of yellow rust, the leaves of plants were grouped into one of 9 classifications of disease incidence (x): 0,1, 10, 20, 30, 45, 60, 80 and 100% covered by rust. 0% represented no incidence of yellow rust, and 100% was the greatest incidence. The disease index (DI) was then calculated using (Li et al. 1989):

$$DI(\%) = \frac{\sum(x \times f)}{n \times \sum f} \times 100$$

where *f* is the total number of leaves of each degree of disease severity and *n* is the degree of disease severity observed (in this work, n ranged from 0 to 8). In each plot, 20 individuals were randomly selected for check.

#### **2.2.1.3 Canopy spectral measurements**

The method of canopy spectral measurements and data was same as the part 1.1.1.2 above.

Crop Disease and Pest Monitoring by Remote Sensing 47

lower-left region (lower range DI).

Xuezao; '□' = 98–100

PRI-derive

d DI (%)

DI (%)

R2 = 0.91


> R2 = 0.97

0 20 40 60 80 100 measured DI (%)

Fig. 5. Comparison of measured DI and PRI-estimated DI for 2003 Exp; 'Δ' = Jing 411;'+' =

regression line (predominantly mid-range DI), and Jing 411 was concentrated in the central

The DI was estimated on a pixel-by-pixel basis in each of the acquired PHI images using the regression equation. To map the degree of yellow rust infection in the trial field, the DI was

**2.2.2.2 Application of multi-temporal PHI images for DI estimation** 

Fig. 4. Plot of measured disease index (DI) as a function of measured photochemical

reflectance index(PRI) for all varieties combined in 2002 Exp. Δ: Jing 411; +: Xuezao; □: 98–100

#### **2.2.1.4 Airborne hyperspectral imaging**

Airborne hyperspectral images of the trial field were acquired in 2003 using the Pushbroom Hyperspectral Imager (PHI) designed by the Chinese Academy of Science (CAS) and flown onboard a Yun-5 aircraft (Shijiazhuang Aircraft Manufacturing Company, China). The PHI comprises a solid state, area array, and silicon CCD device of 780 × 244 elements. It has a field of view of 21o, and is capable of acquiring images of 1 m × 1 m spatial resolution at an altitude of 1000 m above ground. The wavelength range is 400–850 nm with a spectral resolution of 5 nm. Images of the target field were acquired in 2003 at the phenological growth stages of stem elongation (April 18, 2003, Zadoks stage 3), anthesis (May 17, 2003, Zadoks stage 5) and milky maturity (May 31, 2003, Zadoks stage 8). The inoculated wheat was adequately infected by rust on April 18, obviously infected by May 17, and seriously infected by May 31. Measurements of DI were made and in situ canopy reflectance spectra were also acquired on the same dates. All images were geometrically and radiometrically corrected using an array of georeferenced light and dark targets (5 m x 5 m) located at the extremes of the field site. The aforementioned field spectrometer was used to calibrate these targets relative to BaSO4. The location of each target, as well as field measurements of DI were recorded using a differential global positioning system (Trimble Sunnyvale California, USA).

#### **2.2.1.5 Photochemical reflectance index (PRI)**

Because yellow rust epiphyte reduced foliar physiological activity by destroying foliar pigments, the photochemical reflectance index (PRI) was selected as the spectrophotometric method of estimating the disease index. PRI was calculated by the formula in Table 2.

#### **2.2.2 Results**

#### **2.2.2.1 PRI versus DI**

Fig. 4 shows a plot of the measured DI as a function of PRI for all varieties. The data points associated with the variety Xuezao dominate in the top-left region of the scatter plot (relatively high range of DI), while those associated with the variety 98-100 are located in the mid region (mid-range DI) and those associated with Jing 411 dominate the lower right region. This distribution trend is consistent with the relative susceptibility of these varieties to rust; Xuezao is the least resistant and Jing 411 has the greatest resistance. The regression equation of DI using PRI in 2002 Exp was obtained as following (n = 64):

$$\text{DI}(\%) = -721.22 \, (PRI) + 2.40 \qquad \left( -0.14 \le PRI \le 0.02; r^2 = 0.91 \right)$$

An important feature in, the associated regression equation (Fig. 4) was that the spectrallyderived PRI explained 91% of the variance observed in the disease index. This explanation also encompassed the three varieties of wheat as well as the four stages of crop development for each variety. In the subsequent validation of the PRI-DI regression equation with the 2003 Exp data (Fig. 5), the coefficient of determination (R2) between the estimated and measured values was 0.97 (n = 80).

In Fig. 5, the locations of data points associated with individual varieties wew consistent with the levels of resistance to rust. Xuezao dominated the top right-hand region of the scatter plot (relatively high range of DI), the variety 98-100 had points scattered all along the

Airborne hyperspectral images of the trial field were acquired in 2003 using the Pushbroom Hyperspectral Imager (PHI) designed by the Chinese Academy of Science (CAS) and flown onboard a Yun-5 aircraft (Shijiazhuang Aircraft Manufacturing Company, China). The PHI comprises a solid state, area array, and silicon CCD device of 780 × 244 elements. It has a field of view of 21o, and is capable of acquiring images of 1 m × 1 m spatial resolution at an altitude of 1000 m above ground. The wavelength range is 400–850 nm with a spectral resolution of 5 nm. Images of the target field were acquired in 2003 at the phenological growth stages of stem elongation (April 18, 2003, Zadoks stage 3), anthesis (May 17, 2003, Zadoks stage 5) and milky maturity (May 31, 2003, Zadoks stage 8). The inoculated wheat was adequately infected by rust on April 18, obviously infected by May 17, and seriously infected by May 31. Measurements of DI were made and in situ canopy reflectance spectra were also acquired on the same dates. All images were geometrically and radiometrically corrected using an array of georeferenced light and dark targets (5 m x 5 m) located at the extremes of the field site. The aforementioned field spectrometer was used to calibrate these targets relative to BaSO4. The location of each target, as well as field measurements of DI were recorded using a differential

Because yellow rust epiphyte reduced foliar physiological activity by destroying foliar pigments, the photochemical reflectance index (PRI) was selected as the spectrophotometric method of estimating the disease index. PRI was calculated by the formula in Table 2.

Fig. 4 shows a plot of the measured DI as a function of PRI for all varieties. The data points associated with the variety Xuezao dominate in the top-left region of the scatter plot (relatively high range of DI), while those associated with the variety 98-100 are located in the mid region (mid-range DI) and those associated with Jing 411 dominate the lower right region. This distribution trend is consistent with the relative susceptibility of these varieties to rust; Xuezao is the least resistant and Jing 411 has the greatest resistance. The regression

<sup>2</sup> DI % 721.22 2.40 0.14 0.02 0 91 *PRI PRI ; r .*

An important feature in, the associated regression equation (Fig. 4) was that the spectrallyderived PRI explained 91% of the variance observed in the disease index. This explanation also encompassed the three varieties of wheat as well as the four stages of crop development for each variety. In the subsequent validation of the PRI-DI regression equation with the 2003 Exp data (Fig. 5), the coefficient of determination (R2) between the estimated and

In Fig. 5, the locations of data points associated with individual varieties wew consistent with the levels of resistance to rust. Xuezao dominated the top right-hand region of the scatter plot (relatively high range of DI), the variety 98-100 had points scattered all along the

equation of DI using PRI in 2002 Exp was obtained as following (n = 64):

**2.2.1.4 Airborne hyperspectral imaging** 

global positioning system (Trimble Sunnyvale California, USA).

**2.2.1.5 Photochemical reflectance index (PRI)** 

**2.2.2 Results** 

**2.2.2.1 PRI versus DI** 

measured values was 0.97 (n = 80).

Fig. 4. Plot of measured disease index (DI) as a function of measured photochemical reflectance index(PRI) for all varieties combined in 2002 Exp. Δ: Jing 411; +: Xuezao; □: 98–100

Fig. 5. Comparison of measured DI and PRI-estimated DI for 2003 Exp; 'Δ' = Jing 411;'+' = Xuezao; '□' = 98–100

regression line (predominantly mid-range DI), and Jing 411 was concentrated in the central lower-left region (lower range DI).

#### **2.2.2.2 Application of multi-temporal PHI images for DI estimation**

The DI was estimated on a pixel-by-pixel basis in each of the acquired PHI images using the regression equation. To map the degree of yellow rust infection in the trial field, the DI was

Crop Disease and Pest Monitoring by Remote Sensing 49

The results of this work confirm PRI is a potential candidate for monitoring of yellow rust, and could form the basis of an on-the-go sensor and variable-rate spray applicator or remote

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

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

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

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)

Properties of HJ-CCD

(nm) Spatial resolution (m) Swath (km) Revisit time (day)

30 360 2

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

the wide range of wavelengths in multispectral images (Zhang et al., 2011).

SKB is evaluated with both simulated data and field measured data.

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

**2.3.1.3 Acquisition of moderate resolution satellite images** 

**2.2.3 Conclusion** 

detection and mapping process.

**2.3.1 materials and methods** 

processing.

**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** 

Band Wavelength range

Blue 0.430–0.520

Green 0.520–0.600 Red 0.630–0.690 Near-infrared 0.760–0.900

binned into the following classes; very Serious (DI > 80%), serious (45% < DI ≤ 80%), moderate (10% < DI ≤ 45%), slight (1% < DI ≤ 10%) and none (0 < DI ≤ 1%) (Fig. 6).

Fig. 6. Classied DI images derived from PHI airborne images of the trial site in 2003 Exp

Fig. 7 shows the relationship between the DI calculated from the multi-temporal PHI images and the actual measured DI from the 120 sample sites located within the field (R2=0.91).

Fig. 7. Comparison of PHI-derived estimates of DI and actual DI values for 2002 Exp. Data were extracted from all three imaging times, although the DI values were< 20% for the April 18 image

#### **2.2.3 Conclusion**

48 Remote Sensing – Applications

binned into the following classes; very Serious (DI > 80%), serious (45% < DI ≤ 80%),

Fig. 6. Classied DI images derived from PHI airborne images of the trial site in 2003 Exp

PHI-derived DI (%)

18 image

Fig. 7 shows the relationship between the DI calculated from the multi-temporal PHI images and the actual measured DI from the 120 sample sites located within the field (R2=0.91).

> R2 = 0.91

0 20 40 60 80 100 measured DI (%)

Fig. 7. Comparison of PHI-derived estimates of DI and actual DI values for 2002 Exp. Data were extracted from all three imaging times, although the DI values were< 20% for the April

moderate (10% < DI ≤ 45%), slight (1% < DI ≤ 10%) and none (0 < DI ≤ 1%) (Fig. 6).

The results of this work confirm PRI is a potential candidate for monitoring of yellow rust, and could form the basis of an on-the-go sensor and variable-rate spray applicator or remote detection and mapping process.
