(A) high-resolution soil moisture estimation and its application at the regional scale

국문 초록 (Abstract)

토양수분은 물 순환 분석의 핵심 요소로, 강우, 유출, 증발산 사이에 토양에 저장된 수분을 의미한다. 해상도가 낮은 토양수분 자료는 시공간 변동성을 정확히 반영하기 어려우며, 최근 고해상도 SAR 영상의 등장으로 공간적으로 대표성 있는 토양수분 자료 산출을 위한 연구가 활발히 진행되고 있다. 본 연구에서는 기존의 SAR 기반 기법인 WCM(Water Cloud Model) 및 변화탐지(Change Detection)와 함께 딥러닝 기반 기법을 결합한 토양수분 산정 기법을 제안한다.
기존 방법론은 제한된 토양수분 범위(0.1~0.4 m³/m³) 및 식생이 희박한 조건에서 R 값 0.5의 정확도를 보였으며, ANN 기반 기법은 R 값 0.6까지 향상되었으나 낮은 시간 해상도와 불충분한 관측값, 후방산란계수의 지역적 특성 등으로 인해 안정적인 정확도 확보에는 한계가 있었다. 반면, PINN은 다양한 식생 조건과 35–43도의 LIA(Local Incidence Angle) 범위에서도 적용이 가능하며, 최대 0.7, 최소 0.5의 R 값을 확보하여 식생 조건을 극복할 수 있는 가능성을 제시하였다.
또한, AI 기반 컴퓨터 비전 분야에서 도입된 초해상화(Super-Resolution) 기법을 통해 기존 위성자료의 공간 해상도를 개선하고, SAR 영상의 낮은 시간 해상도 문제를 보완할 수 있음을 확인하였다. SMAPL4 자료를 대상으로 수행된 실험에서는 약 3배의 공간 해상도 확대에도 불구하고 평균 R 값 0.5 수준의 정확도를 확보하였다. 보조자료 없이 기존 자료의 해상도를 개선하거나, 새로운 고해상도 토양수분 자료를 생성하는 기술은 위성 기반 토양수분 활용의 확장에 있어 중요한 연구결과로 판단된다.
지형이 복잡하고 기상 조건이 다양한 한반도에서는 시공간 해상도가 개선된 토양수분 자료가 높은 활용 가능성을 지니며, 본 연구는 새로운 기법을 통해 안정적인 고해상도 토양수분 자료 산출과 공간 기반 토양수분 정보의 활용도 확장을 위한 기반을 마련하고자 한다.

다국어 초록 (Multilingual Abstract)

Soil moisture is a key factor for water cycle analysis and refers to the moisture stored in the soil between rainfall, runoff, and evapotranspiration.
This means that it is difficult to accurately capture soil moisture variability with coarse resolution soil moisture, and with the recent advent of high-resolution images; SAR, research is being conducted to produce high-spatial-resolution soil moisture data to increase the usability of spatially representative soil moisture data.
In this study, a soil moisture retrieval technique using deep learning techniques, as well as WCM and Change detection, which are basic SAR-based soil moisture production techniques are suggested.
The existing methodology showed an R value of 0.5, which is calculated for limited soil moisture conditions (0.1 to 0.4 m3/m3) and sparse vegetation conditions. In the case of ANN, an accuracy of R value of 0.6 was shown, but it was difficult to obtain stable accuracy due to the low temporal resolution, insufficient observation values, and the behavior characteristics of the backscattering coefficient that strongly showed regional characteristics. For PINN, an accuracy of up to 0.7 and a minimum of 0.5 are secured for various Local Incidence Angles ranging from 35 degrees to 43 degrees, and vegetation conditions are applied without restrictions, suggesting a method to overcome vegetation conditions. In addition, the applicability of the super resolution technique recently introduced in AI-based computer vision to improve the spatial resolution of existing satellite data to overcome the temporal resolution of SAR images are confirmed. The validation results conducted on SMAPL4 confirm a technique that produces images with a spatial resolution magnification of 3 times with an average accuracy of 0.5. The introduction of a new high-resolution soil moisture data generation technique and a technique for improving the resolution of existing data without auxiliary data is considered a meaningful research result in a situation where satellite data is expected to expand in the future. In the case of Korea, where there are complex terrains and diverse meteorological conditions, it is expected that soil moisture data with improved spatiotemporal resolution will be highly useful. This study is conducted for the purpose of retrieving stable high-resolution soil moisture data using a new technique, thereby establishing a foundation for expanding the usability of spatial soil moisture data.

목차 (Table of Contents)

• 1. INTRODUCTION 1
• 1.1 Background 1
• 1.1.1 In-situ soil moisture 1
• 1.1.2 Remote sensing soil moisture 2
• 1.1.3 High resolution soil moisture 8
• 1.2 Motivation 13
• 1.2.1 Interference of vegetation on SAR signal 13
• 1.2.2 Application of super-resolution technique on existing soil moisture product 15
• 1.3 Objective of the research 16
• 2. STUDY AREA AND DATASET 20
• 2.1 Study area 20
• 2.2 Dataset 25
• 2.2.1 Sentinel-1 SAR 25
• 2.2.2 In-situ soil moisture dataset 28
• 2.2.3 SMAPL4 soil moisture product 29
• 3. METHODOLOGY 30
• 3.1 Preprocessing of SAR data 30
• 3.2 Radar Vegetation Index (RVI) 35
• 3.3 SAR based soil moisture retrieval 38
• 3.4 Super resolution technique for soil moisture product downscaling 47
• 4. RESULTS 49
• 4.1 Relation analysis between soil moisture and backscattering for adopting physical methods 49
• 4.2 Comparison of linear regression and WCM on 10 m scale 52
• 4.3 Analysis of WCM results for vegetation type 57
• 4.3.1 Results of WCM-based soil moisture according to vegetation conditions 60
• 4.3.2 Error analysis according to the range of observed soil moisture values for each vegetation condition 64
• 4.4 Deep learning based SAR soil moisture retrieval 66
• 4.4.1 Analysis of forest utilization of SAR VV backscatter coefficients using ANN technique 66
• 4.4.2 Independent utility analysis of SAR soil moisture 72
• 4.5 Development of PINN model and evaluation 73
• 4.6 Application of super-resolution technique for soil moisture downscaling 103
• 5. CONCLUSION 109
• References 113