**2.2 Field preparation and data acquisition**

The experimental site is located in Pathumthani Province, Thailand (14° 12' N, 100° 37' E) for the study. The site may be located anywhere in the world, but, the soil properties have to be measured accurately. The soil of the experimental site belonged to the clay textural class with average bulk density of 1.38 g cm-3 and pH of 4.2. Three replicates were made, and the treatment plots, each of size 10 m 10 m, randomly distributed within each replicate. To estimate the nitrogen application rate, the total nitrogen present in the soil was tested using standard methods (Kjeldahl apparatus). At the experimental site, the concentration of preexisting nitrogen was classified as low (<0.18%) level for all the plots, as per the local Agricultural Extension Service guidelines.

For the underlined experiment, the plots were well-watered using flood irrigation and carefully maintained for pest control to ensure uniform yield potential. The rice seeds were broadcasted (on 14 Dec. 2006) in accordance with local practices under irrigated farming conditions. Nitrogen fertilizer was applied at five rates: 0%, 25%, 50%, 75%, and 100% of recommended values, representing 0, 33, 66, 99, and 132 kg ha-1, respectively. Plots with different nitrogen rates were maintained to promote a wide range of rice yield so the effectiveness of LARS images could be evaluated for varying nutrient availability. This follows a similar technique by Chen et al. (2003), who used four N rates (0, 45, 90, and 135 kg ha-1) in field experiments with a Tainung 67 rice crop for multispectral image analysis. An early rice variety, Supanburi-1 (95 day period), was used in the study, as this is one of the most popular variety in central Thailand. Urea (46-0-0) was applied as the source of nitrogen for the study. Different nitrogen rates along with recommended phosphorous fertilizer were applied 30 days after sowing rice. Images were obtained twice with the LARS system just before panicle initiation stage (45 and 65 days after planting, Figure 1b).

The altitude has to be selected considering the camera's field of view to acquire a single image for each treatment plot. Images with effective dimensions of 18 m 14 m were collected from a 20 m flying height, covering a single plot. For field application the height can be varied as per the suitability of the researcher, to cover wider area in each image. Flight altitude was recorded with a height sensor (MPXAZ4115A barometric sensor, Freescale Semiconductor, Austin, Tex.) mounted on the LARS system. Images are obtained at five different heights, and the images obtained closest to the 20 m height were selected for analysis. Five ground-based reflectance readings were obtained for the rice canopy and BaSO4 standard white reference board using the *Spectroradiometer* in each of the experimental plots. The ground-based readings were obtained immediately after the LARS system-based image acquisition. The plot-wise ground-based reflectance value is calculated as the mean of the five readings.

#### **2.3 Image pre-processing**

258 Remote Sensing of Biomass – Principles and Applications

As individual images, of digital camera covers very small ground area, is mosaic with the algorithm to develop a single map of the study area. HIPSC software converts the digital numbers into relevant sensor readings and used them to carry out image processing operations, such as: image rotation, image mosaic and reflectance index (NDVI, Green NDVI etc.) estimation. The software can develop site-specific zone maps based on variation in reflectance index values and also provide ground control points (GCPs) for mosaic image

The experimental site is located in Pathumthani Province, Thailand (14° 12' N, 100° 37' E) for the study. The site may be located anywhere in the world, but, the soil properties have to be measured accurately. The soil of the experimental site belonged to the clay textural class with average bulk density of 1.38 g cm-3 and pH of 4.2. Three replicates were made, and the treatment plots, each of size 10 m 10 m, randomly distributed within each replicate. To estimate the nitrogen application rate, the total nitrogen present in the soil was tested using standard methods (Kjeldahl apparatus). At the experimental site, the concentration of preexisting nitrogen was classified as low (<0.18%) level for all the plots, as per the local

For the underlined experiment, the plots were well-watered using flood irrigation and carefully maintained for pest control to ensure uniform yield potential. The rice seeds were broadcasted (on 14 Dec. 2006) in accordance with local practices under irrigated farming conditions. Nitrogen fertilizer was applied at five rates: 0%, 25%, 50%, 75%, and 100% of recommended values, representing 0, 33, 66, 99, and 132 kg ha-1, respectively. Plots with different nitrogen rates were maintained to promote a wide range of rice yield so the effectiveness of LARS images could be evaluated for varying nutrient availability. This follows a similar technique by Chen et al. (2003), who used four N rates (0, 45, 90, and 135 kg ha-1) in field experiments with a Tainung 67 rice crop for multispectral image analysis. An early rice variety, Supanburi-1 (95 day period), was used in the study, as this is one of the most popular variety in central Thailand. Urea (46-0-0) was applied as the source of nitrogen for the study. Different nitrogen rates along with recommended phosphorous fertilizer were applied 30 days after sowing rice. Images were obtained twice with the LARS system just before panicle initiation stage (45 and 65 days after

Fig. 2. Schematic representation of the LARS image acquisition system

geo-registration using commercial software.

**2.2 Field preparation and data acquisition** 

Agricultural Extension Service guidelines.

planting, Figure 1b).

Multispectral images acquired by the Tetracam ADC camera (.dcm format) were converted into *.tiff* format for analysis. The *.tiff* format reduces the storage space and also effectively retains the image quality for image processing. Images were uploaded to Pixelwrench software (Tetracam, Inc., Chatsworth, Cal.), which contains programs for deriving one of several vegetation indices (.hdr format) from raw image data. An NDVI image was produced for each test plot, and the average NDVI index was estimated using the customdeveloped program in the C programming language from images acquired by the LARSmounted sensors (Figure 3). Ground-based reflectance data were collected to estimate mean NDVI of the experimental plots (NDVISPECTRO). NDVISPECTRO was estimated using the software provided by the *Spectroradiometer* manufacturer. Linear regression models can be developed in SAS (ver. 9.1, SAS Institute, Inc., Cary, N.C.) or any standard software.

Fig. 3. Stages of image processing: (a) raw image with plot boundaries (as taken by the image acquisition system), (b) plot-scale image of the rice crop, and (c) NDVI image

#### **3. Validation of LARS setup**

The normalized difference vegetation index (NDVI) is the mostly adopted reflectance index for agricultural cropping and vegetation studies (Rouse et al., 1973) given as;

$$NNDVI = \frac{NIR - R}{NIR + R} \tag{1}$$

Where, *NIR*: Radiance value for *Near-infrared* band; *R*: Radiance value for *Red* band.

Rice Crop Monitoring with Unmanned Helicopter Remote Sensing Images 261

Fig. 4. Variation of vegetation index with N-treatment rates; a) NDVISPECTRO; b) NDVILARS ; c)

GNDVILARS

The Green Normalized Differential Vegetation Index (GNDVI) to establish the suitability of reflectance index for rice cropping with variable nitrogen rates (Gitelson et al., 1996) was also used. GNDVI, based on the greenness level, represented by the chlorophyll content determining the radiance level of the leaf surface, was very significant for the rice crop monitoring. The GNDVI was estimated as follows,

$$\text{GNDVI} = \frac{\text{NIR} - \text{G}}{\text{NIR} + \text{G}} \tag{2}$$

Where, *NIR*: Radiance value for *Near-infrared* band; *G*: Radiance value for *Green* band. The NDVI index was also calculated from ground level *Spectrophotometer* radiance values using the Eqn. 1 for establishing suitability of LARS system. Around five readings were taken from each plot in order to estimate the average NDVI for each treatment plots. The *SPAD 502 meter* readings of leaf greenness can be converted into Chlorophyll content by the following equation for rice cropping (Markwell et al., 1995).

#### **3.1 Relationship between reflectance indices and variable N-treatments**

The graph of NDVISPECTRO and NDVILARS plotted for the different N-treatments showed positive response with increased recommended nitrogen rates. The NDVI index, taken 45 days after sowing, showed weak relationship with nitrogen treatment rates, attaining coefficient of determination (r2) of 0.60. As, the fertilizer application, just two weeks before date of testing, response time may not be enough to influence plant leaf radiance level to greater extent (Figure 4). However, the relationship was stronger (r2 ≈ 0.85) with higher NDVI values, ranging from 0.70 to 0.90, for second set of *Spectrophotometer* reading taken at booting stage (for 65 days old plants). NDVILARS, estimated from LARS images, were very low, after 45 days of sowing, ranging from 0.2 to 0.6, due to the lower radiance value of soil, exposed in gaps between the plants' leaves. The radiance level of the crop leaves, covering the whole plot area with least exposed area at booting stage, attained their original values (with NDVI between 0.85 to 1.0). NDVILARS at booting stage showed strong relationship with r2 of 0.73 for different N-treatment rates (Figure 4b). The greenness index (GNDVI) plotted against variable nitrogen rates showed, lower correlation with r2 of 0.66 and 0.7, for the images taken at 45 days and 65 days respectively, with slightly strong relationship for the later. The lower range of GNDVILARS index had values ranging from 0.5 to 0.6 at booting stage, maintained positive response with higher nitrogen rates (Figure 4c).

#### **3.2 Suitability of reflectance indices determined from LARS images**

Cross comparison analysis was carried out to testify the applicability of LARS images through indices such as NDVILARS and GNDVILARS with the *Spectrophotometer* reading index such as NDVISPECTRO by plotting graphs between them (Figure 5). The NDVILARS was proportional to that of NDVISPECTRO with r2 of 0.72 and 0.79 for 45 days and 65 days old rice crop, respectively. The NDVILARS ranges from 0.85 to 1.0 showing sound crop coverage throughout the plot at booting stage of crop. The lower range NDVILARS value (0.2~ 0.5) for 45 days crop made the reading unsuitable to represent the crop in crop modeling and predictions. The higher r2 value (≈ 0.7) for indices estimated from LARS images (NDVILARS, GNDVILARS) with index from ground *Spectrophotometer* reading (NDVISPECTRO) showed the suitability of the proposed system for crop status studies.

The Green Normalized Differential Vegetation Index (GNDVI) to establish the suitability of reflectance index for rice cropping with variable nitrogen rates (Gitelson et al., 1996) was also used. GNDVI, based on the greenness level, represented by the chlorophyll content determining the radiance level of the leaf surface, was very significant for the rice crop

*NIR G GNDVI*

The graph of NDVISPECTRO and NDVILARS plotted for the different N-treatments showed positive response with increased recommended nitrogen rates. The NDVI index, taken 45 days after sowing, showed weak relationship with nitrogen treatment rates, attaining coefficient of determination (r2) of 0.60. As, the fertilizer application, just two weeks before date of testing, response time may not be enough to influence plant leaf radiance level to greater extent (Figure 4). However, the relationship was stronger (r2 ≈ 0.85) with higher NDVI values, ranging from 0.70 to 0.90, for second set of *Spectrophotometer* reading taken at booting stage (for 65 days old plants). NDVILARS, estimated from LARS images, were very low, after 45 days of sowing, ranging from 0.2 to 0.6, due to the lower radiance value of soil, exposed in gaps between the plants' leaves. The radiance level of the crop leaves, covering the whole plot area with least exposed area at booting stage, attained their original values (with NDVI between 0.85 to 1.0). NDVILARS at booting stage showed strong relationship with r2 of 0.73 for different N-treatment rates (Figure 4b). The greenness index (GNDVI) plotted against variable nitrogen rates showed, lower correlation with r2 of 0.66 and 0.7, for the images taken at 45 days and 65 days respectively, with slightly strong relationship for the later. The lower range of GNDVILARS index had values ranging from 0.5 to 0.6 at booting stage, maintained positive response

Where, *NIR*: Radiance value for *Near-infrared* band; *G*: Radiance value for *Green* band. The NDVI index was also calculated from ground level *Spectrophotometer* radiance values using the Eqn. 1 for establishing suitability of LARS system. Around five readings were taken from each plot in order to estimate the average NDVI for each treatment plots. The *SPAD 502 meter* readings of leaf greenness can be converted into Chlorophyll content by the

**3.1 Relationship between reflectance indices and variable N-treatments** 

**3.2 Suitability of reflectance indices determined from LARS images** 

suitability of the proposed system for crop status studies.

Cross comparison analysis was carried out to testify the applicability of LARS images through indices such as NDVILARS and GNDVILARS with the *Spectrophotometer* reading index such as NDVISPECTRO by plotting graphs between them (Figure 5). The NDVILARS was proportional to that of NDVISPECTRO with r2 of 0.72 and 0.79 for 45 days and 65 days old rice crop, respectively. The NDVILARS ranges from 0.85 to 1.0 showing sound crop coverage throughout the plot at booting stage of crop. The lower range NDVILARS value (0.2~ 0.5) for 45 days crop made the reading unsuitable to represent the crop in crop modeling and predictions. The higher r2 value (≈ 0.7) for indices estimated from LARS images (NDVILARS, GNDVILARS) with index from ground *Spectrophotometer* reading (NDVISPECTRO) showed the

*NIR G* 

(2)

monitoring. The GNDVI was estimated as follows,

following equation for rice cropping (Markwell et al., 1995).

with higher nitrogen rates (Figure 4c).

Fig. 4. Variation of vegetation index with N-treatment rates; a) NDVISPECTRO; b) NDVILARS ; c) GNDVILARS

Rice Crop Monitoring with Unmanned Helicopter Remote Sensing Images 263

Biomass is a plant attribute that is time consuming and difficult to measure or estimate, but easy to interpret. Biomass is regarded as an important indicator of ecological and management processes in the vegetation. Biomass estimation facilitates accurate management decisions regarding chemical and fertilizer applications, estimation of yield, and post harvest handling of stover (Pordesimo et al., 2004).Quantifying spatial variation in pasture and crop biomass can help to direct management practices and improve farm productivity, through accurate and informed movements of grazing rotations, crop and pasture nutrient management and also yield prediction (Trotter et al., 2008). Measurement of plant biomass by harvesting is destructive, expensive and time consuming (Reese et al., 1980). de Matthaeis et al. (1995) used AIRSAR data collected over the agricultural fields to monitor biomass variation. They found that the L-band is more effective for crops with low

The rice biomass (threshed rice plant without the grain) of three sampled areas, 4 m2 each, were collected and weighted. The moisture content (w.b.) of the threshed rice plant was estimated using standard method. The dry weight of the threshed rice plant was estimated

Total oven-dried (Abdullah et al., 1992) biomass was ranged from 3.58 to 7.36 ton ha-1 for the different treatments (Table 4). Total dry biomass weight between the treatments showed significant differences at the 0.10 level but no significant difference between replicates.

0 kg ha-1 3.58 4.25 6.30 4.710 33 kg ha-1 5.51 5.84 5.64 5.660 66 kg ha-1 5.57 5.97 5.77 5.771 99 kg ha-1 6.50 7.36 5.97 6.611 132 kg ha-1 5.57 6.63 7.30 6.501

Linear calibrations curves were developed in SAS 9.1 to estimate the biomass from NDVI index values calculated from LARS images. From these results, NDVILARS could explain 76%

of the variation in biomass weight (r2 = 0.760, RMSE = 0.598 ton ha-1, Figure 6).

Total biomass (ton/ha) = <sup>10000</sup> (100 . .) . 12 1000 *M C BiomassWt* (3)

Replicate Average 1 2 3

plant density, while C-band is better for high plant density crops.

and converted into the total biomass weight per ha i.e. (ton/ha).

M.C.: Moisture content of weighed rice plant (w.b.)

Table 3. Total biomass (ton ha-1) of the experimental plots

Total biomass: Weight of rice plant (without rice grain) in ton/ha BiomassWt: Weight of threshed rice plant (without rice grain)

**4. Estimation of crop parameters** 

**4.1 Estimation of total biomass** 

**4.1.1 Calculations** 

Where,

N Rate Treatment

Fig. 5. Comparison of indices based on groundtruthing data and LARS images; NDVISPECTRO with; a) NDVILARS; b) GNDVILARS

#### **3.3 Discussions**

For the experiment, the recommended amount of fertilizer was applied to 40 days old crop and the first set images and groundtruthing were taken at 45 days. The leaf coverage was low with a major share of exposed soil resulting in lower correlation of green indices (NDVI and GNDVI) values. The coefficient of determination (r2) was improved visibly for 65 days old crop with denser crop leaving little exposed soil. As observed, 65 days old crop, LARS and ground measurements, was better suited, hence selected for crop status monitoring studies. Variation in green indices (NDVI and GNDVI) showed symmetry with the variation of nitrogen level for different treatments.
