Utilizing Large Language Models for enhanced soil moisture prediction and gap-filling in satellite-derived data

Soil moisture (SM) is a critical variable in land-atmosphere interactions, playing a significant role in numerous environmental and climatic processes. Continuous and long-term monitoring of SM is vital for various applications, including agriculture, hydrology, and climate change assessment. Satellite missions have made substantial contributions to SM monitoring. However, the continuity and completeness of satellite-derived SM time series are often compromised by data gaps. These gaps are caused by factors such as satellite revisit intervals, and radio frequency interference (RFI) contamination, which pose weak supervisions for traditional machine learning models from the data perspective. In recent years, pretrained foundational models, exemplified by Large Language Models (LLMs), have demonstrated remarkable reasoning capabilities in data scarcity scenarios. However, the linguistic nature of LLMs makes it challenging to directly process complex numerical information. To align the modalities of geography and natural language, we carefully design a set prompts to related variables into LLMs. These geophysical and hydrometeorological variables include land surface temperature, precipitation, vegetation index, potential evaporation, soil texture, geographical coordinates, and land cover type. By incorporating these variables, the model can better capture the complex interactions and dependencies influencing SM dynamics. The methodology involves applying the different techniques in both temporal and spatial domains and rigorously evaluating their performance using the holdout cross-validation technique. The integration of various data sources and the application of advanced modeling techniques aim to enhance the accuracy and reliability of soil moisture predictions in gap areas.The findings of this study are expected to contribute significantly to the field of remote sensing and SM monitoring by providing a robust framework for gap-filling in satellite-derived datasets.