Application of Deep Learning for an Improved Satellite-based Soil Moisture Retrieval Algorithm

Abstract

Soil moisture (SM) is one of the important variables in understanding the hydrological cycle on the Earth's surface. The characterization of SM was conducted using in situ, land surface modeling, and satellite data; these three sources were used complementarily. The satellite SM provides a spatio-temporally continuous observation-based SM globally. The parameterization of the early stage of the SM retrieval algorithm is conducted in a limited space and time. Therefore, existing algorithms could be biased to some extreme earth conditions that were underrepresented in the parameterization stage. This dissertation tried to improve existing satellite SM retrieval algorithms by integrating various sources of SM data with deep learning (DL) algorithms. Chapter One provides an introduction to available SM data, satellite-based SM data, and DL approaches. Chapter Two tries to seek a potential strategy to apply DL to SM estimation. The study demonstrated that the minimization of scale-mismatch between grid-scale SM data and point-scale in-situ SM data can make DL-based SM estimation successful. By considering lessons from the second chapter, Chapter Three tries to improve existing satellite-based SM retrieval algorithms using DL and examines the potential way to improve SM retrieval algorithms. The results from Chapter Three showed that the existing retrieval algorithm’s RTM parameters need to be adjusted. Chapter Four seeks a new way to fully parametrize the SM retrieval algorithm by combining physics equations with DL. The results of Chapter four showed potential for capturing SM variations over densely vegetated regions from the current generation microwave satellite mission. Chapter Five summarizes the work shown in this thesis. The experiments shown in this dissertation might help us provide valuable satellite observation-based SM data to understand various eco-hydrological phenomena.