A Deep Learning Framework for Soil Moisture Retrieval with Sentinel-1 Short Time Series

Abstract

Soil moisture refers to the quantity of water present within the unsaturated zone of the soil. It is a key indicator of many Earth’s surface processes. Accurate and timely knowledge of its spatial and temporal distribution is indispensable for many applications such as drought early‑warning systems or flood forecasting. Current Synthetic Aperture Radar (SAR) systems such as the Sentinel-1 constellation have attracted the attention of the scientific community to improve the quality and the resolution of soil moisture products. In our study, we investigate a novel deep learning-based (DL) solution for accurate and time-tagged soil moisture retrieval by combining, for the first time, backscatter and repeat-pass interferometric information derived from Sentinel-1 multi-temporal data. To overcome the challenge posed by the scarcity of high-quality reference data required for fully-supervised training, we propose a two-step method: a weakly-supervised pre-training of the model on a larger amount of data with lower accuracy, followed by a fully-supervised fine-tuning, from on-ground high-reliable measurements of soil moisture from 0 to 5 cm depth.