HydroChronos: Forecasting Decades of Surface Water Change

TL;DR

HydroChronos delivers the first large-scale, multi-modal dataset for forecasting surface water dynamics, integrating decades of Landsat-5 and Sentinel-2 imagery with climate and DEM data across Europe, the US, and Brazil. It defines three standardized forecasting tasks and introduces AquaClimaTempo UNet (ACTU), a climate-aware spatiotemporal baseline that leverages a dedicated climate encoder and gated fusion, supported by regression losses combining multiscale and wavelet components. Empirical results show ACTU consistently outperforms persistence across change detection, direction classification, and magnitude regression, with ablations clarifying the contributions of input modalities and loss design. An Explainable AI framework—including climate-subgroup discovery, global feature attribution, and per-channel saliency—reveals key climate drivers (e.g., precipitation, soil moisture, temperature) and spectral-band importance, offering actionable guidance for robust, climate-aware hydrological forecasting and future dataset development.

Abstract

Forecasting surface water dynamics is crucial for water resource management and climate change adaptation. However, the field lacks comprehensive datasets and standardized benchmarks. In this paper, we introduce HydroChronos, a large-scale, multi-modal spatiotemporal dataset for surface water dynamics forecasting designed to address this gap. We couple the dataset with three forecasting tasks. The dataset includes over three decades of aligned Landsat 5 and Sentinel-2 imagery, climate data, and Digital Elevation Models for diverse lakes and rivers across Europe, North America, and South America. We also propose AquaClimaTempo UNet, a novel spatiotemporal architecture with a dedicated climate data branch, as a strong benchmark baseline. Our model significantly outperforms a Persistence baseline for forecasting future water dynamics by +14% and +11% F1 across change detection and direction of change classification tasks, and by +0.1 MAE on the magnitude of change regression. Finally, we conduct an Explainable AI analysis to identify the key climate variables and input channels that influence surface water change, providing insights to inform and guide future modeling efforts.

Figures (6)

• Figure 1: Distribution of lakes and rivers in HydroChronos
• Figure 2: Sample in the three modalities: optical (RGB channels only for visualization), DEM, and climate.
• Figure 3: RGB sample at two different timesteps and the corresponding MNDWIs.
• Figure 4: Visual example of tasks for Lake Tahoe. In regression, the values range from 0 to 2 (blue to red). In change detection, labels are no-change (blue) and change (red). In direction classification, labels are negative change (blue), no-change (grey), and positive change (red).
• Figure 5: AquaClimaTempo UNet (ACTU) architecture. If DEM is provided, it is repeated once per sample in the image timeseries and concatenated along the channel axis. The Pyramidal Image Feature Extractor provides multiscale embeddings. If a climate timeseries is provided, the climate encoder provides multiscale embeddings which are gate fused with the image embeddings. ConvLSTMs provide multiscale embeddings for the timeseries, which are used in the UNet decoder to provide the final prediction.