Hyperspectral remote sensing of rice agriculture for field scale variability mapping

It is now believed that mapping the biophysical and biochemical parameters, such as chlorophyll, nitrogen and leaf water content, to determine field-scale heterogeneity for a rice farming system is feasible using hyperspectral remote sensing. The central objective of the present study is to investigate the possibility of optimal solutions involving hyperspectral remote sensing images for mapping paddy crop parameters at plot scale within the crop fields.

Figure 1.1 : Potential of remote sensing in the field of agricultural studies (Panda et al., 2010)

Precision rice farming: Global Developments

Precision farming system (PFS) aims to increase agricultural crop productivity with minimal environmental impact through proper management of farm variability. In Germany, a group of researchers has worked to integrate both card and sensor approaches with variable rate technologies (VRT) for chemical applications on nitrogen fertilizer applicators (O'Neal et al., 2000).

Figure 2.1 : Integrated system for information management of rice precision farming (Source:

Nutrient management in paddy cropping systems

Previous studies have investigated the capability of remote sensing platforms against various agricultural practices and management systems in different parts of the world. Rapid development of optical remote sensing system methods has provided a number of practical means for agricultural applications.

Figure 2.2 : The visible spectrum for monitoring plant health (Campbell, 1996) Green

Optical remote sensing

He found that the Normalized Difference Vegetation Index (NDVI), which generally absorbed high red reflectance, was used to determine biophysical parameters to identify the state of pure vegetation. It revealed that MODIS-derived indices such as Normalized Difference Vegetation Index (NDVI) and Normalized Difference Water Index (NDWI) using SWIR (1640 nm or 2130 nm) are successful in estimating VWC.

Advanced remote sensing technologies

Microwave remote sensing applications for crop parameters studies

Vegetative water content in sunflower can be determined using X- and C-band data (Santi et al., 2012). Therefore, the average relative water content of the canopy can be quantified by the leaf water content index (Hunt et al., 1987).

Table 2.3 : Estimation of crop chlorophyll content using hyperspectral data (Space-borne measurement)

Study Site 1: RRLRRS in Kamrup district of Assam

Areas with hot and humid climate accompanied by heavy rainfall are mostly suitable for rice cultivation. For the present study, two study sites were considered, one in Kamrup district of Assam (North East India) and the other in Cuttack district of Odisha (East India).

Study Site 2: Farmers’ agricultural fields in Cuttack district of Odisha

The tremendous diversity in growing conditions makes the classification and characterization of rice environments a challenging task. To understand rice plant characteristics, crop parameter variability, crop variety information, fertilizer application and irrigation scheduling, field experiments were conducted in rice fields.

Figure 3.2 : Photograph showing the paddy crop with different nitrogen applications

Ground-based hyperspectral data collection

The spectral ranges of the sensors together with spectroradiometer were 350-1050 nm with a field of view of 25°. For each observation, the spectral reflectance values were averaged as the final spectral reflectance representing each rice field of respective rice varieties on a specific date of observation (Figure 3.5).

Table 3.1 : Field Observation of paddy crop in the year 2009 and 2014 Rice

Field survey

Another field campaign was conducted on farmers' agricultural fields in Cuttack, Odisha (study site 2) during the pre-monsoon period of the year 2016 (Figure 3.8). Satellite data from different sensors were collected to study the field-scale variability of paddy crops.

Figure 3.6 : Photograph showing farmer interaction for winter rice during the field campaign

Hyperspectral Satellite Data

Hyperspectral data pre-processing

The radiometrically corrected Hyperion L1R contains vertical lines in different individual bands of the data set. Further, in order to reduce the data dimensionality of the dataset, the minimum noise fraction transform (MNFT) was adopted to separate the bands dominated by noise from the bands containing meaningful information.

Table 3.2 : Hyperion scene characteristics of Study site 1 in Assam during monsoon period

Multispectral Satellite Data

Currently, in relation to high spatial and spectral resolutions, hyperspectral satellite remote sensing has been used as an effective method in estimating the concentration of vegetation nutrients in a large cultivated area (Darvishzadeh et al., 2008). In addition, hyperspectral remote sensing of biophysical and biochemical parameters has been achieved through the development of new hyperspectral indices (Galvao et al., 2009; Thenkabail et al., 2000).

Table 3.4 : LISS IV sensor data specification Spectral Band

Study area

Data used

By knowing these critical wavelengths, one will be able to easily analyze the effect of nitrogen on different rice genotypes. Stepwise procedure has been described in the flowchart (Figure 4.3) to distinguish between different nitrogen treatments using improved vegetation indices.

Figure 4.1 : The field layout of experimental plots

Vegetation Indices

From the Figure 4.14, it can be observed that using three nitrogen-based vegetation indices, growth occurred in an increasing trend, peaked and then decreased towards the end of the crop development age indicating the senescence period. From Figure 4.16 it is clearly reflected that the crop is under dry zone in its early crop stage.

Figure 4.4 : Variation of nitrogen treatments at 67 DAP and 82 DAP

Statistical Analysis

Then, the irrigation water was supplied to the crops in the middle of the development stage of the crop. The variance analysis of the spectral reflectance of rice varieties using the water index is presented in Table 4.4 and showed satisfactory results for the studied rice genotypes.

Table 4.3 : Analysis of variance of the spectral reflectance of rice varieties using the studied indices

Varietal effects

It is very difficult to distinguish the rice crop species with the naked eye in their early stages as they look very similar. Rice species discrimination with different nitrogen treatments and extraction of pure rice spectra for the development of rice crop spectral library is presented.

Table 5.1 : Details of ground based hyperspectral measurements

Waveform classification approach

Experimental measurements have been assessed to characterize canopy reflectance properties in a rice farming system accompanied by different rice species and nitrogen treatments.

Cluster analysis of rice varieties

The clustering method basically follows a distance measurement to find similarity or dissimilarity between pairs of entities. The distance between the two entities can be measured using the Minkowski metric (Han and Kamber, 2011) and is given by .

Significant Wavebands

The methodology proposed in this study performed well in differentiating rice cultivars with different nitrogen treatments. A waveform classification approach was adopted in this study to build a spectral library by eliminating noisy waveforms present in spectral acquisition.

Spectral reflectance characteristics

Here it is observed that the spectral response of all the varieties is similar in nature at 67 DAP. From Figure 5.4, the spectral signatures found are quite different for some of the variants at 82 DAP.

Discrimination of rice varieties

Following this hierarchical agglomerative clustering technique, twenty-four rice varieties were clustered and formed into different groups (Figure 5.7). In addition, some of the studied varieties were grouped, resulting in single variety or mixed rice varieties.

Table 5.2 reveals that some of the rice varieties got separated from each other, whereas some of the varieties were not distinguished

Critical wavebands

These significant wavelengths were consistent with the findings (810,830 nm) obtained by Inoue et al. 2008) sensitive to nitrogen. Furthermore, leaf nitrogen content is a strong factor influencing both optimal light use efficiency and canopy photosynthesis rate (Kergoat et al., 2008).

Table 5.4 : Waveband combinations having least correlation between narrow bands (N 3 -N 1 ) Name of the variety

Ground-based hyperspectral measurements

Hyperspectral measurement from EO-1 satellite

Spectral indices are the true indicators of the biophysical parameters of agricultural crop (Chen et al., 2010; Heiskanen et al., 2013). Here x and y are the reflectances (Ri and Rj) or first derivative (Di and Dj) values at i and j nm over the entire hyperspectra (Inoue et al., 2008).

Estimation of Chlorophyll content

Although the indices, OSAVI and MTCI, provided significant results for rice chlorophyll content, they had negligible results in the estimation of wheat chlorophyll content (Bannari et al., 2008). In addition, the performance of the modified indices (mSR and OSAVI) was found to be remarkable for the assessment of rice chlorophyll content.

Table 6.4 : Best correlated chlorophyll indices for the estimation of paddy crop chlorophyll content

Estimation of Nitrogen content

The standard error associated with predicted values of leaf nitrogen content of less than 3% is estimated as per Tian et al. 2011) and present SR models modified by statistical formula and found to be 46% and 33% respectively. From Figure 6.2, it is observed that the present and Tian et al. 2011) LNC model perform opposite to each other in predicting nitrogen content of rice.

Table 6.5 : Relationship between the nitrogen indices calculated from ground based hyperspectral spectra and nitrogen content

Agricultural pigment and nutrient mapping from hyperspectral imagery

Application of published indices to map nitrogen content of rice

Similar to Figure 6.5 (a, b), the result shows that the non-linear OSAVI model overpredicted the chlorophyll content of rice up to 9 mg/g, while the linear model estimated it to be 3.5 mg/g. In Figure 7.6 (f), the spatial distribution obtained from the NDII linear regression model showed a good acceptable performance, where the LRWC varied widely from 62 to 92%. It matches well with the water content of rice leaves because the 1600 nm water absorption band is highly correlated. on the water content of the plant.

Table 6.6 : Nitrogen content statistics derived from N classified map

Application of published indices to map chlorophyll content of rice 119

Clustering Analysis of paddy crop

Waveforms were extracted from Hyperion images for each of these sites and waveform classification followed by hierarchical clustering analysis was performed. From Figure 6.9, it is clearly observed that by using hierarchical clustering technique, different groups of rice genotypes are reformed and embedded noise is also captured.

Figure 6.10 : Clustering of spectral waveforms from Hyperion image, 3 rd October, 2014 (Rice agriculture site-II)

Spatial Distribution of rice agriculture system

Within the observed chlorophyll range obtained from the studied rice varieties grown in the rice farming system, these were the index models that well predicted the chlorophyll content of paddy crops based on Hyperion images. The heterogeneity of the crop in the field can be easily quantified from hyperspectral satellite images and hence the spatial variation of Chl and N mapping of paddy crop based on Hyperion images can be used to provide an informative system for accurate to develop rice farming system.

Figure 6.12 : Spatial distribution of rice agriculture site - II

In-situ hyperspectral data

Rice varieties like Jaya, Abhishek, Chandrama, Shabhagi Dhan, Ranjit, Baismuthi, Nilanjana, IR64, Joymati, Vandana, Kolong and Naveen were cultivated. There was provision of irrigation for the water needs of the plants, to provide sufficient water for the rice plants throughout the growing period of the plants.

Space-borne hyperspectral data

Narrowband indices for estimating leaf water content can be expressed as Normalized Difference Water Index (NDWI) and Simple Ratio Water Index (SRWI). Mathematically, it can be defined as, y). In addition, the SWIR vegetation water indices were tested to map the leaf water content of the paddy in a crop calendar during the pre-monsoon and monsoon periods.

Figure 7.1 : Complete sequence to map LRWC variability from space platformFind the correlation between

Estimation of Leaf Relative Water Content

Determination of critical stage for LRWC measurement

This model cannot provide significant field-level variability for the winter rice farming system. These index models were derived specifically for green crops in the Indian rice farming system from hyperspectral data.

Figure 7.2 : Scatter plot between WBI and LRWC

Development of SWIR water index models from VNIR water index models147

Mapping of LRWC for winter rice agriculture system

The LRWC determined from Hyperion images ranged from 68% to 90%, which is in good agreement with the observed LRWC of the studied rice genotypes which ranged from 61% to 92% for the winter rice farming system. Figure 7.6 (d) shows that the NMDI model, which showed satisfactory results from in-situ observations, underestimates the LRWC for winter rice agriculture based on Hyperion images ranging from 64% to 72%.

Mapping of LRWC for summer rice agriculture system

Furthermore, the LRWC variation of the summer rice farming system using the MSI linear regression model derived from NWI-4 from the Hyperion image ranged from 65 - 95% (Figure 7.7 (c)). On the other hand, interesting observations resulted from the NMDI regression model (Figure 7.7 (d)) on the LRWC map in a summer rice farming system.

Figure 7.7 : Variability mapping of LRWC in a summer rice agriculture system derived from different narrow band index models during pre-monsoon period ((a): NDWI-1, (b): NDWI-2, (c): MSI, (d): NMDI, (e): SRWI, (f): NDII, (g): SAWI-1, (h): SAWI-2, (i): WI

Derivation of multispectral bands (Synthetic bands) from Hyperion data

Band Average Concept

NCare the critical narrow bands of green, red and NIR regions in the spectrum respectively. RIH, NIHas the three broad bands corresponding to integrated narrow bands in green, red and NIR regions of the spectrum respectively.

Figure 8.1 : Correlation between critical bands and broad bands equivalent to integrated narrow bands, green: (a)-(b), red: (c)-(d), NIR: (e)-(f)

Spectral Shape Function Concept

For this LISS IV image representing the study area dominated by rice crop cultivation, a linear relationship was established between multispectral broadband (so-called synthetic band; .. SLISS) and broadband equivalent to integrated narrow bands of the Hyperion image (IH ). Secondly, it is a difficult task to find the field-level parameters of rice crop, such as chlorophyll and nitrogen, from a space platform.

Figure 8.2 : Band to band correlation analysis to get the critical bands from spectral shape function method, (a) green, (b) red, (c) NIR

Performance of band average method

For the present study, estimation of paddy crop parameters from LISS IV images (of the Study Site 2) acquired on 22 February 2016 was carried out using narrowband index regression models. Therefore, it has been specially achieved to extract the rice crop variability such as chlorophyll and nitrogen at the plot level within the agricultural field.