Adaptive Residual-Informed Deep Transfer Learning for Accurate Flood Depth Estimation Using Spatiotemporal Feature Learning

Prediction of the depth of floods is crucial in minimizing the devastating impact of floods on human settlement, structures and environment based on timely risk assessment and competent management of the calamity. Despite all these advances, many models face issues in an accurate pixel based flood depth estimation, particularly in the urban and coastal complexes and there is need to develop a predictive model that is more robust and efficient. To address them, this paper proposes a new hybrid framework that integrates two more recent approaches: SPYRA model of spatiotemporal residual feature extraction and MIRAI network of MoCo-informed residual adaptive inference. The SPYRA model indicates the hierarchical and spatial-temporal pattern of flood images and the MIRAI network improves the situation by the use of adaptive learning and contrastive transfer learning, which enables predicting the depth of floods to be valid and explainable. The suggested work has been unique because it integrates the residual features extraction and adaptive inference synergistically hence enhancing the generalization and reducing the prediction error and efficiency of the computational techniques. The experiment results demonstrate the usefulness of the proposed framework over nine additional deep learning and machine learning frameworks. MIRAI model has the highest accuracy of 99%, lowest RMSE of 0.13, MAE of 0.09, highest R2 of 0.94 and lowest MAPE of 3.2% which is the power and effectiveness of the model. The additional tests including a feature distribution, error distributions, and aligning the actual and the predicted depth of the floods also confirm the validity and the interpretability of the proposed method.