How deep-learning modules match ecological time scales for smarter water forecasts. (IMAGE)
Caption
How deep-learning modules match ecological time scales for smarter water forecasts. This study compared nine time-series forecasting architectures across two reservoirs with contrasting trophic states (mesotrophic Königshütte with 3 predictors, and eutrophic Yazidang with 10 predictors) to predict chlorophyll a—a key proxy for algal blooms. The graphical summary reveals that forecast performance is not model-agnostic but is governed by specific core modules aligned with ecological rhythms. Patch embedding acts as a universal temporal operator, consistently boosting accuracy by up to 10.1%. Channel-independent processing (e.g., in PatchTST) excels in short-term (1–7 days) forecasting (Nash-Sutcliffe efficiency up to 0.88) by capturing the strong self-persistence of algal biomass. In contrast, cross-variable attention (e.g., in Crossformer) proves superior for medium-term (8–15 days) predictions (efficiency up to 0.83) by untangling delayed nutrient-temperature interactions. Explainable AI further confirms a temporal shift in predictor reliance—from current chlorophyll a to temperature and nutrients as lead times extend. These architectural-ecological alignments establish a blueprint for building interpretable, horizon-adaptive AI systems for environmental monitoring.
Credit
Environmental Science and Ecotechnology
Usage Restrictions
Credit must be given to the creator. Only noncommercial uses of the work are permitted. No derivatives or adaptations of the work are permitted.
License
CC BY-NC-ND