Hydrological Applications with Multi-Sensor Satellite Data and AI: Inundation Detection and Water Quality Prediction Seongjun Lee College of Engineering Gwangju Institute of Science and Technology

Abstract

Climate change has intensified hydrological variability, exacerbating the frequency of extreme flood events and the degradation of inland water quality (WQ). While remote sensing offers a viable pathway for monitoring these phenomena, traditional approaches struggle with limitations such as cloud cover, complex scattering mechanisms, and, most critically, the lack of ground-truth data in ungauged or inaccessible regions. This thesis addresses these challenges by integrating multi-sensor satellite data with advanced Artificial Intelligence (AI) techniques through two distinct hydrological applications: flood inundation detection and WQ prediction. In the first part, we developed a robust approach for high-resolution flood mapping in data-denied regions using Polarimetric Synthetic Aperture Radar (PolSAR). To overcome the spatial transferability issues of convolutional neural network (CNN)-based deep learning (DL) models, we proposed a metaheuristic optimization (MO)-based feature selection method. By employing a swarm-based simulated annealing (SwarmSA) algorithm, we identified an optimal subset of features invariant to domain shifts, including VV polarization and Gray-Level Co-occurrence Matrix (GLCM) textures. The proposed approach, utilizing a Gated Shape CNN architecture, was trained on a flood event in Iran (arid climate) and tested on an unseen event in North Korea (monsoon climate). The results demonstrated superior generalization capability (IoU = 0.8863, F1-score = 0.9397), confirming that physics-informed feature engineering enables reliable monitoring in inaccessible areas without local retraining. In the second part, we tackled the data scarcity problem in inland WQ prediction using optical satellite data. We established a fine-tuning approach to predict WQ parameters in small-scale reservoirs where in-situ data is insufficient for training DL models from scratch. Utilizing Harmonized Landsat and Sentinel-2 (HLS) data and ERA5-Land reanalysis data, a multi-layer perceptron (MLP) model was pre-trained on the three major river basins of South Korea and then fine-tuned for small target dams (Gwangju Dam and Seomjingang Dam). The fine-tuning strategy successfully overcame the domain shift, improving the coefficient of determination from negative values (model collapse) to over 0.60 for chlorophyll-a. Furthermore, the integration of Shapley additive explanations-based analysis validated the physical reliance on red-edge bands, and Monte Carlo dropout provided crucial uncertainty quantification. Collectively, this thesis demonstrates that the synergy of multi-sensor satellite data and AI, specifically through generalized feature selection and domain adaptation, provides a scalable and scientifically robust solution for hydrological monitoring in data-limited environments.