Trustworthy Hydro-ML Benchmark Framework

This repository contains the framework and code for benchmarking rainfall-runoff models, evaluating them across three critical dimensions: predictive accuracy, interpretability, and physical plausibility. It includes four distinct machine learning architectures, specifically the tree-based Random Forest (RF) and XGBoost (XGB) models, and the sequence-based Temporal Convolutional Networks (TCN) and Long Short-Term Memory (LSTM) networks. These are benchmarked against the Hydrologiska Byråns Vattenbalansavdelning (HBV) model, a conceptual process-based hydrological model, and the HYdrologic Prediction with multi-model Ensemble and Reservoir computing (HYPER) model, a hybrid framework combining multi-model ensembles with reservoir computing. Prediction accuracy is assessed by comparing the observed and simulated river discharge using the Kling-Gupta Efficiency (KGE) metric. Interpretability is evaluated using SHapley Additive exPlanations (SHAP) analysis to diagnose the internal logic and feature importance of the models. For physical plausibility, the models are subjected to three diagnostic stress tests: mass conservation is assessed by evaluating long-term water balance and evapotranspiration consistency; drought extrapolation is tested by simulating recession to zero discharge under synthetic zero-precipitation scenarios; and flood extrapolation is evaluated by testing the physical realism of peak discharge magnitude and timing under synthetic extreme rainfall events with increasing intensity and duration.