LAUSR

ABSTRACT Accurate vegetation moisture estimation is essential for wildfire prediction, climate modeling, and ecosystem monitoring. However, traditional ground-based methods lack scalability and fail to provide continuous spatial coverage, presenting a critical research gap in large-scale vegetation monitoring. To address this, we introduce SARMoistX, an explainable and scalable framework that integrates SAR C-band backscatter with weather data for temporal vegetation moisture prediction and leveraging eXplainable AI (XAI) techniques to interpret model decisions. Our methodology evaluates multiple machine learning models, including xLSTM, TFT, and Mamba, using robust validation strategies and key performance metrics such as MAE, RMSE, and MAPE. Results indicate that xLSTM provides the best balance between accuracy and computational efficiency, while TFT achieves slightly higher accuracy at a significantly greater computational cost. An ablation study highlights that Backscatter VV polarization outperforms VH for vegetation moisture prediction, with a hybrid VV-VH approach yielding minimal improvement. Additionally, Monte Carlo dropout-based uncertainty quantification was employed to assess model reliability, further enhancing the interpretability of SARMoistX. The framework’s scalability was tested across different vegetation types, demonstrating its adaptability to diverse ecological conditions. These findings underscore the potential of SAR-based temporal vegetation moisture estimation as a robust and scalable solution, advancing remote sensing applications in environmental monitoring and resource management.