Abstract
Purpose Soil Water Content (SWC) is a critical factor in precision agriculture, influencing crop health, irrigation planning, and land management. Current methods for estimating SWC have low spatial resolution, fail to account for the spatial impact of surrounding areas, and are often limited to surface SWC. Therefore, this research aims to assess the feasibility of a high-resolution, multi-depth (root zone) SWC estimation method tailored for precision agriculture applications. Methods We propose a deep learning-based approach that integrates remote sensing and multimodal data to estimate SWC at high spatial resolution across multiple depths. Our method combines a U-Net model for spatial feature extraction, a Temporal Convolutional Network (TCN) for time-series processing, and a Feed-Forward Neural Network (FNN) for contextual information. A key challenge in this task is the scarcity of ground truth data due to the limited number of in-situ SWC measurements. To address this, we introduce the Relative Soil Water Content (RSWC) parameter, which enhances surface SWC estimation by leveraging historical remote sensing data. Results Using two field cases, we evaluate our model against two state-of-the-art methods: a point-based deep learning model and a numerical model. Results demonstrate that our approach outperforms both baselines in SWC estimation across different depths, achieving Mean Square Errors (MSEs) of 1.54% and 2.01% for the two fields, compared to 2.69% and 3.37% for the point-based method and 3.82% and 6.21% for the numerical model. Conclusions Our method generates high-resolution, multi-depth SWC maps for the entire field without requiring extensive in-situ measurements, presenting a multimodal deep learning approach as a practical proof-of-concept solution for large-scale agricultural applications.