Overcoming data scarcity: A transfer learning framework with fine-tuned neural networks and multi-sensor satellite image fusion for soil moisture estimation

Training deep learning (DL) models requires extensive data, particularly for soil moisture prediction, where large volumes of in situ measurements are needed to prevent overfitting. To address this challenge, we propose a customised transfer learning framework that adapts a pre-trained DL model to a new study site with a different climate. Specifically, we fine-tune a fully connected feed-forward neural network, originally trained on a large dataset from a humid subtropical region (source domain), using limited data from a semi-arid region (target domain). The proposed framework leverages nine input features extracted and generated from Sentinel-1/2 and Shuttle Radar Topographic Mission (SRTM) images through a linear data fusion technique. We systematically evaluate the performance of the proposed framework against ten benchmark algorithms. We observed that the proposed framework outperforms all benchmark algorithms, achieving a correlation coefficient (R) of 0.81, a root mean square error (RMSE) of 0.05 $m^3/m^3$, and a bias of 0.02 $m^3/m^3$ on the target domain. Particularly, this is achieved using 55% less in situ data compared to the source domain. To ensure reliability and robustness, we conduct comprehensive analyses, including error histogram, residual, uncertainty, spatial distribution, ablation, statistical, and complex time complexity analyses. Throughout each evaluation, the proposed framework consistently exhibits a reliable and robust performance. The findings of this study hold promise in facilitating accurate surface soil moisture estimation, particularly in data-scarce regions.