A scoping review of spatiotemporal ConvLSTM applications for predicting water balance components

Seyed Hossein Hosseini, Harri Koivusalo, Jussi Nikander, Henrikki Tenkanen

Deep learning is renewing computational hydrology by offering advanced capabilities for modeling complex environmental processes characterized by spatiotemporal variability. Among these approaches, the Convolutional Long Short-Term Memory (ConvLSTM) network has gained considerable attention for its ability to learn spatial and temporal dependencies simultaneously, a feature particularly valuable for accurately predicting water balance components. Such predictions are essential for understanding hydrological processes, guiding water resource management, and informing climate adaptation strategies. In this work, we systematically combine scattered information about studies that have applied ConvLSTM to predict water balance–related variables, synthesize their methodologies and findings, and identify key research gaps to guide future developments. Our review shows that ConvLSTM possesses a flexible architecture and can be equipped with attention mechanisms, encoder–decoder structures, and deformable convolutions to improve its predictive accuracy for hydrometeorological variables. Beyond spatiotemporal prediction, reported applications include multisource satellite data fusion, bias correction, and spatial downscaling. Finally, we outline future research directions, including integrating physical constraints into ConvLSTM architecture, developing hydrologically meaningful explainable artificial intelligence methods, advancing spatiotemporal uncertainty quantification, and coupling ConvLSTM with available hydrological models.