XAI-driven analysis of influencing factors in streamflow-based flood forecasting via fine tuned MTV19Anet

An accurate and timely flood prediction that relies on streamflow information is important for developing an early warning system to minimize the negative effects of floods. This paper proposes a new Steam flow-based Flood prediction model using a Fine-tuned MTV19ANet architecture (SFFPM-FTA). The novelty of proposed model integrates an Enhanced Normalization Technique (enNormT) that effectively reduces the impact of large outliers. It also employs an improved entropy-based feature extraction method to enhance prediction robustness. The hyperparameters are optimized using an Improved Wolverine Optimization Algorithm (IWoOA) with SHAP, an explainable AI technique to insights into model interpretability. The study also includes the visual representation of steam flow, where normal flow and flooded flow are visually represented over river areas. The experimental analysis shows that the proposed SFFPM-FTA model obtains high performance compared to other prediction models with respect to various error measures. The proposed model demonstrates minimal errors in streamflow-based flood forecasting through improved normalization during preprocessing and enhanced entropy-based feature extraction. Compared with other baseline methods, the proposed method demonstrates superior performance, with reductions of 55.26% in MAE, 64.56% in MAPE, 83.45% in MSE, and 57.30% in RMSE, respectively. In addition, the model achieves NSE of 0.988, KGE of 0.963, and R² of 0.986, demonstrating its effectiveness in accurately capturing streamflow dynamics. Overall, the proposed SFFPM-FTA improves flood forecasting accuracy and provides reliable and interpretable streamflow predictions.