A Multi-Modal Spatio-Temporal Dataset for Land Subsidence Forecasting from InSAR Time-Series

Advancements in deep learning for spatio-temporal forecasting are often hindered by the complexity and domain-specific nature of raw geoscience data. To bridge this gap, we present a novel, analysis-ready dataset for multi-modal land subsidence forecasting, derived from the European Ground Motion Service (EGMS) InSAR time-series. This dataset is meticulously pre-processed and structured to facilitate the development and benchmarking of sophisticated AI models. It moves beyond traditional uni-modal approaches by fusing three critical types of information into a unified data cube: (1) Dynamic Features, representing the raw displacement time-series; (2) Static Physical Priors, including key geophysical attributes like mean velocity, acceleration, and seasonality; and (3) Cyclical Temporal Features, which encode the day-of-year to capture periodic patterns. All features are rasterized onto a 64x64 grid, making the dataset directly compatible with modern deep learning frameworks. By providing this rich, multi-modal representation, our dataset aims to foster a new paradigm of physics-informed machine learning research in the geophysical sciences.