DEVELOPMENT OF MACHINE LEARNING AND ANALYTICAL APPROACH FOR SOIL MOISTURE RETRIEVAL WITH LOW AND HIGH RESOLUTION SATELLITE DATA

Abstract

Over the last decade with the advancement of space technology, a variety of sensors (optical, thermal, and microwave) have been deployed on spaceborne, airborne, and UAV (unmanned aerial vehicles) platforms to acquire multitudes of information from the earth's surface. Sometimes single sensor does not suffice the information required for an application due to its technical limitations. These limitations may result in poor spatial or temporal resolution or can be affected by weather conditions. For an application like soil moisture (SM) estimation from the vegetative field, it is required to have high temporal and good spatial resolution with precise information on fraction vegetation cover. This challenge may be conquered by using data fusion techniques with the efficient application of electromagnetic modeling (EM) and machine learning (ML). However, each sensor's band properties, temporal and spatial resolution may differ, making it a challenging task to synergistically use them. Hence, for optimum utilization of these sensors, information fusion with analytical and ML approaches has become an emerging field of research that finds its applications in land cover classification, crop classification, fraction vegetation cover (FVC), soil moisture estimation, etc. The main emphasis of this thesis is to study and develop methods for FVC estimation using optical and SAR features with help of ML, EM for total backscattering computation from heterogeneous vegetation field and coarse resolution pixel, and segregation of soil contribution from it for soil moisture estimation, further development of information fusion approach for the resolution enhancement. The general objective is supported by performing the following tasks: (i) Study and development of a machine learning-based model for FVC estimation using optical and SAR data. (ii) (iii) (iv) Development of an FVC-based soil moisture retrieval algorithm for Sentinel-1 data. Development of soil moisture estimation algorithm for coarse resolution Scatsat-1 data with synergetic use of urban and vegetation fraction cover. An information fusion approach for disaggregation of coarse resolution Sigma0 and soil moisture product. The thesis is organized into seven chapters, which are briefly summarized below. Chapter 1 presents the introduction, motivation, scope, and objectives of the current research work. A brief literature review related to the tasks performed in the thesis is given in Chapter 2. The first objective emphasizes the study and development of a machine learning-based model for fraction vegetation cover (FVC) estimation using optical and SAR data, which is given in Chapter-3. FVC is a ratio of the vertical projection area of green vegetation to the total area under consideration. Empirical models, linear un-mixing models, and physical-based models are widely used by researchers for FVC estimation using optical data. The limitation of these models is given below: ❖ The empirical model is limited for the specific study sites for which it is developed. ❖ The challenges arise in finding pure endmembers of different classes in linear unmixing models, e.g., the dimidiate pixel model. ❖ Physical-based models are complex and require various input parameters. ❖ Optical data is not available in the cloudy season. Concerning the above problems, various machine learning (ML) models have been used for FVC estimation using features derived from optical and synthetic aperture radar (SAR) data as input features. The novelty and contribution made by this chapter are given below: ❖ To overcome the weather dependency of optical data in this chapter, an in-depth analysis of SAR features is presented. Entropy, alpha, VH, VH/VV, and VV (power) are the optimum SAR features for FVC estimation. ❖ It is also observed that using optimum features combination provides exemplary results during the entire crop cycle for FVC estimation compared to a single feature. These features can be used as an alternative for FVC estimation when optical data is unavailable. ❖ Xgboost, light GBM, and random forest give more effective results for FVC estimation. ❖ This chapter overall novelty is exploring the SAR data for FVC estimation using ML. In the second objective, an FVC-based soil moisture retrieval algorithm for Sentinel-1 data has been proposed and given in Chapter-4. This chapter investigates an electromagnetic-based model for soil moisture (SM) estimation in crop cover areas. The most cumbersome application is to estimate the SM from the crop-covered soil; due to the complexities involved in the segregation of soil and vegetation contribution from the total backscattering signal. In the literature review, several EM, such as polarimetric decomposition, change detection, and water cloud model (WCM) been proposed to deal with the complexities involved in the segregation of soil and vegetation contribution. The problems for application of these models is given below: ❖ Polarimetric decomposition requires fully polarimetric SAR data. ❖ The change detection approach requires high temporal time series data. ❖ Researchers widely prefer WCM, but it was developed for homogeneous vegetation layers and requires optical or ground truth data for vegetation information. ❖ Algorithms are weather-dependent due to the use of optical data-based vegetation information. Concerning the above problems, WCM has been modified by including the FVC to model the backscattering signal from the heterogeneous vegetation field. The proposed modified water cloud model (MWCM) segregates the soil contribution from the total backscattering signal, which is further used in the modified Dubois model for SM estimation. The novelty and contribution made by this chapter are given below: ❖ MWCM gives more accurate soil and vegetation contribution modelling from the heterogeneous vegetation field, hence improving the SM accuracy. ❖ A constrained nonlinear multivariable optimization technique is used for selecting the optimum value of model parameters. ❖ Sentinel-1 SAR-derived vegetation descriptors have been used; hence the developed model has the all-weather working capability and single sensor dependent. Chapter 5, "Synergetic use of urban and vegetation fraction cover for SM estimation using coarse resolution Scatsat-1 data," presents a detailed analysis of SM estimation using coarse resolution satellite data. In this work, data obtained from the Scatsat-1 scatterometer has been used. This research work tries to answer the following questions: ❖ How to precisely model the backscattered signal for the coarse resolution because, in the coarse resolution, more than one classes are the significant contributor? ❖ What is the feasibility of Scatsat-1 scatterometer data for the SM application because it is primarily launched for ocean and weather applications? The WCM is modified by including the fraction cover of different classes to model the backscattering signal from the coarse-resolution pixel, and to segregate the soil contribution from coarse resolution pixel. Which is further used in the Dubois model and Topp's model for SM estimation. The novelty and contribution made by this chapter is given below: ❖ For the first time in literature, urban fraction cover has been considered to model the total backscattering signal and its segregated soil contribution. In the coarse resolution pixel, urban becomes a significant contributor, and it is necessary to consider the urban effect. ❖ It is observed that the segregated soil backscattering temporal contribution pattern is in line with the temporal rainfall pattern. Hence it proved that SM estimation is feasible with the Scatsat-1 data. ❖ A critical analysis has been performed to observe the urban effect on the estimated SM, and it is found that the estimated SM is overestimated without urban correction. v ❖ The retrieved SM is compared with the rainfall data and ground truth SM, and it is observed that the variation in SM is in line with the rainfall occurrence over the study area and very close to the ground truth SM. Chapter 6 proposes the methodology: "Developing an information fusion approach for disaggregation of coarse resolution backscattering signal and SM." In this work, freely available multisensor data has been used. SM estimated with scatterometer data (σ°) has a high temporal resolution, but its spatial resolution limits its applications for regional scale. Hence to make it suitable for regional applications, resolution enhancement becomes essential. This can be achieved by either enhancing the σ° product itself or developing a model for the resolution enhancement of SM. This research work tries to answer the following questions: ❖ What is the feasibility of downscaling the coarse-resolution backscattering signal by utilizing the freely available high-resolution data, which has not been explored much yet? ❖ How to make the SM downscaling algorithms weather independent? Concerning the above problems, a multisensor information fusion-based approach has been proposed to downscale the coarse resolution backscattering signal and SM. The novelty and contribution made by this chapter are given below: ❖ An adaptive random weighing variance-based approach is proposed to downscale the backscattering signal, which selects the optimum fusion factors by the variance minimization method for each parameter to obtain a weight factor image. ❖ This weight factor image is used to downscale the coarse-resolution backscattering signal. A methodology is also proposed to downscale the coarse-resolution SM to high-resolution SM with the spatio-temporal information fusion of Sentinel-1 SAR data. Required auxiliary information to downscale the SM is computed with the sentinel-1 data, which makes the algorithms weather-independent Finally, Chapter 7, summarizes the obtained results and enlists the significant contributions made to the thesis. The perspective of the future scope of this work is also discussed.