Enhancing Soil Moisture Prediction with Data-Driven Models: A Global Perspective

The prediction of soil moisture plays a vital role in assessing water availability, optimizing agricultural resources, and preparing for climate-induced disasters. However, significant gaps remain in soil moisture observation networks due to data sparsity, inconsistent temporal coverage, and limited spatial resolution, particularly in underrepresented regions. The International Soil Moisture Network (ISMN), the largest archive of in situ soil moisture data, highlights these challenges, with many datasets averaging only a decade of temporal coverage and biased spatial distribution heavily skewed toward the Global North. This study presents a data-driven modeling framework designed to enhance soil moisture prediction by leveraging advanced machine learning techniques, diverse geospatial datasets, and in situ observations.

Our multi-stream model integrates high-resolution data from Sentinel-2 (NDVI, B4, B8), ECMWF weather forecasts, and SRTM elevation models to predict surface and rootzone soil moisture at six-hour intervals. Validation against SMAP L4 datasets demonstrates high accuracy, achieving mean RMSE values of 0.1087 m³/m³ for surface moisture and 0.1183 m³/m³ for rootzone moisture across 20 Köppen-Geiger climate zones. The modular design enables the model to adapt to diverse climatic conditions and refine predictions through continuous validation. Performance analysis reveals strong temporal generalization and superior results in wet climates, though arid and extreme environments pose challenges, highlighting areas for targeted improvements.

To address data sparsity, the study emphasizes balanced sampling and the integration of citizen science initiatives, which supplement traditional networks by providing localized, high-frequency observations. By incorporating in situ ISMN datasets, the framework aligns with the session's focus on improving observation networks and leveraging data quality assurance. Additionally, hybrid approaches that combine physical constraints with machine learning models ensure predictions are grounded in realistic soil behavior and spatial consistency.

This research underscores the importance of sustained investment in developing and maintaining soil moisture observation networks, particularly in underrepresented regions. It highlights the need for standardized data collection protocols, advanced calibration techniques, and open-access platforms that integrate in situ and satellite observations. By bridging gaps in traditional networks, the model advances global soil moisture monitoring, supporting applications in sustainable agriculture, water resource planning, and climate resilience.

Aligned with session HS8.3.2, this study exemplifies the role of innovative measurement techniques and data-driven approaches in enhancing the utility of soil moisture datasets. The findings advocate for a collaborative scientific effort to address the pressing challenges of data availability, quality assurance, and network deployment. Through scalable modeling frameworks, this research sets the foundation for predictive systems that provide actionable insights to policymakers and practitioners in hydrology, agriculture, and climate science.