| dc.description.abstract | The implementation of sustainable agricultural, hydrological, and environmental management entails an improved understanding of soil properties and its conditions at increasingly finer resolutions. In this context, the detailed reflectance spectra provided by hyperspectral sensors can be beneficially employed in acquiring information about the topsoil. The recent availability of hyperspectral data from newly launched sensors and the advancements in techniques for handling and analyzing hyperspectral images have given rise to a wealth of new research findings in soil monitoring and mapping. Though hyperspectral imaging spectroscopy has facilitated the mapping of soil properties at large scales with finer resolutions, it is limited to only bare soil pixels. This is because the presence of non-soil cover in the form of photosynthetic or non-photosynthetic vegetation in a pixel affects the reflectance spectra by causing a “spectral mixing effect”, which in turn would affect the performance of soil property estimation models. Hence it is essential to identify the bare soil pixels in a study area and address the issue of spectral mixtures prior to soil property mapping. In this context, the reported study aimed at addressing this issue of spectral mixtures to improve the knowledge base on topsoil properties, particularly clay content, in a study area characterized by heterogeneity in terms of soil type and fertility, area under cultivation, crop type and cropping system. This work focused on the following three objectives:
To analyze the effect of different spectral reduction strategies prior to endmember extraction in a linear spectral mixture model in terms of pixel reconstruction errors.
To propose a thresholding approach for bare soil identification using pixel soil fractions obtained from spectral unmixing prior to clay content mapping and compare it with the classic method of using spectral indices
To develop a novel soil fraction-based composite mapping approach to extend the spatial coverage of predicted clay maps.
Spectral mixture modelling is one of the most important techniques for classifying hyperspectral data at sub-pixel resolution and identifying spectrally pure endmembers for estimating their corresponding abundances is an important step in spectral unmixing. The application of spectral reduction techniques prior to endmember extraction for unmixing would optimize the process by increasing the sensitivity of the algorithms to the most distinctive and informative features of the dataset. The first part of the study compared different spectral reduction techniques prior to endmember extraction on six real hyperspectral datasets, including an Airborne Visible InfraRed Imaging Spectrometer-Next Generation (AVIRIS-NG) image over Indian sub-continent. Spectral reduction of the datasets were applied using both feature extraction techniques like Principal Component Analysis (PCA), Independent Component Analysis (ICA), Minimum Noise Fraction (MNF), and a feature selection technique based on Partial Informational Correlation (PIC) measure, along with a no reduction case where all the spectral bands were used. A PIC based spectral reduction was employed for endmember extraction specifically in the context of spectral unmixing for the first time in this study. The locations of the endmembers were identified from these reduced datasets using four endmember extraction algorithms- Pixel Purity Index (PPI), N-FINDR, Automatic Target Generation Process (ATGP) and Vertex Component Analysis (VCA). The endmembers identified from the different combinations of spectral reduction and endmember extraction techniques were used for linear spectral unmixing of the original datasets. The performance of each of the combinations after unmixing were compared in terms of the pixel reconstruction errors and the computation times for each dataset. It was observed that the PIC based spectral reduction performed well in terms of reconstruction errors and computation times when combined with the N-FINDR endmember extraction algorithm. This approach could be recommended for spectral reduction in unmixing of datasets from heterogeneous study areas with similar endmember classes.
The second part of the study aimed to analyze the impact of bare soil pixel identification on clay content estimation using an airborne hyperspectral image. Two methods were tested for identifying the bare soil pixels (i) using a combination of two spectral indices, Normalized Difference Vegetation Index (NDVI) for identifying photosynthetic vegetation and Cellulose Absorption Index (CAI) for identifying non-photosynthetic vegetation and (ii) using soil fraction obtained from spectral unmixing using three endmembers: soil, photosynthetic and non-photosynthetic vegetation. The study used an AVIRIS-NG image and laboratory measured clay content of 272 soil samples acquired over an area of 300 sq. km, in the Gundlupet taluk, Karnataka, India. Two sets of bare soil pixels were identified using the two methods and partial least squares regression (PLSR) models were calibrated and validated to estimate the clay contents over the study area from each. The performances of the regression models and predicted clay content maps were analyzed and compared. It was observed that the PLSR model based on bare soil pixels identified by unmixing provided better performances (R2 of 0.61) than the one using spectral indices (R2 of 0.46) for validation, even though the area mapped was reduced by half (14.96% of the study area) as compared to the latter (30.31%). This study highlighted that an improvement in prediction performance comes at the cost of reduction in spatial coverage in mapping of clay content. Finally, the study also brought forth the need for studying other spectral perturbing factors such as rugosity (due to ploughing) which may explain the modest performances of the clay prediction models, even using a hyperspectral image characterized by high spatial and spectral resolutions.
The limited utility of imaging spectroscopy for topsoil property mapping requires approaches for maximizing the bare soil coverage so as to extend the mapped area. In this regard, a novel soil fraction-based composite mapping approach for clay content estimation from a single AVIRIS-NG image over heterogenous crop fields was proposed as the last part of this study. The approach classified the image pixels according to their soil fraction determined by spectral unmixing and assigned specific regression models to each soil fraction class to estimate clay content along with the prediction uncertainty. The composite clay map gave modest performances with R_val^2 ranging from 0.53 to 0.63 for soil fraction thresholds varying from >0.3 to >0.7 respectively and showed correct spatial pattern irrespective of the soil fraction classes. In addition, the effect of the soil fraction threshold on clay content estimation was analyzed by comparing the performances of regression models built using bare soil pixels identified from varying soil fraction thresholds. It was observed that the model performance increases with adoption of higher soil fraction thresholds in terms of an increase in R_val^2 and decrease in RMSEP values. The potential mapped area in terms of clay content in this study ranged from 10.39% to 52.81% with a soil fraction threshold of >0.7 and >0.3 respectively, and so the compositing approach allowed an extension of the mapped surface by 42.42%. The proposed approach could be adopted to extend the mapping capability of planned and current hyperspectral satellite missions where a combination of the proposed soil fraction-based thresholding and multi-temporal image compositing could significantly increase both the extent of the mapped area and associated prediction performances.
The proposed studies on combination of spectral reduction and endmember extraction techniques prior to unmixing, the bare soil identification analyses and soil fraction-based composite clay mapping approach are envisaged as a premise for future researchers to develop on. The scope of its applicability may be broadened and refined with the inclusion of more sensors, soil properties and analysis techniques. The proposed framework may thus be extended to incorporate newer study areas. | en_US |
