Spatially explicit deep learning analysis of climatic and biophysical controls on long-term Ethiopian lake-area dynamics

Understanding how spatial and temporal factors influence lake areas is crucial for effective water resource management, particularly in regions with limited hydrological records. We developed a spatially explicit deep learning framework for six major lakes in Ethiopia, using long-term sub-catchment climate and biophysical variables as predictor variables, and remotely sensed lake area as response variables. The framework used an Average Hop Downstream Diffusion graph convolution to represent upstream-to-downstream sub-catchment connectivity, followed by average (AVG) and attention (ATTN) pooling to derive basin-level spatial embeddings at each time step. These embeddings were processed with Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) networks, yielding four model variants: AVGGRU, ATTNGRU, AVGLSTM, and ATTNLSTM. All four variants showed strong predictive skill, with correlation coefficients and coefficients of determination above 0.85 and absolute percentage errors below 1 percent of long-term mean lake area, but limited ability to capture high-frequency variability highlights. Model differences were more evident in reproducing variability in amplitude and temporal phase, where skill was more moderate. ATTNGRU showed the most stable overall performance. Attribution analysis based on learned attention weights and gradient sensitivity revealed physically consistent patterns. In large inflow-dominated lakes, the most influential sub-catchments were located in upstream tributary sources; in smaller lakes they were concentrated near shorelines. Rainfall and vegetation were more influential in upland catchments, while temperature, evapotranspiration, and runoff were more important near the lakes. Overall, connectivity-aware attention-based models improved both prediction and process interpretation of lake area dynamics in ungauged, data-scarce systems. Future research on data-driven lake-area prediction is recommended to capture high-frequency variability.