HydroChronos: Forecasting long-term surface water dynamics from multi-modal data

Rege Cambrin, D., Poeta, E., Pastor, E., Corley, I., Cerquitelli, T., Baralis, E., & Garza, P.
ACM Transactions on Spatial Algorithms and Systems
2026

Forecasting surface water dynamics is essential for water resource management and climate change adaptation, but progress is limited by the scarcity of large-scale benchmarks and by the lack of insight into how multi-modal models combine satellite, climate, and topographic signals. In this paper, we provide a detailed characterization of HydroChronos, a large-scale multi-modal spatiotemporal dataset for surface water forecasting that spans over three decades of aligned Landsat 5 and Sentinel-2 imagery, climate variables, and Digital Elevation Models across lakes and rivers in Europe, North America, and South America. We also formalize three forecasting tasks and develop AquaClimaTempo UNet, a spatio-temporal model with a dedicated climate branch and gated fusion mechanism. Across the proposed benchmarks, our model outperforms a persistence baseline by +14% and +11% F1 on change detection and direction-of-change classification, and by 0.1 MAE on magnitude regression. In addition, we present an extended Explainable AI analysis that highlights the most influential climate variables and input channels, offering new insight into the drivers of surface water change.

[Rege Cambrin et al., 2026] Rege Cambrin, D., Poeta, E., Pastor, E., Corley, I., Cerquitelli, T., Baralis, E., & Garza, P. (in press). HydroChronos: Forecasting long-term surface water dynamics from multi-modal data. ACM Transactions on Spatial Algorithms and Systems. Accepted for publication.