Game-Theoretic Consensus Deep Learning for Adaptive Flood Prediction in Digital Twin Environments

We propose Game-Theoretic Consensus Deep Learning (GTCDL), an adaptive AI framework designed to enhance disaster prediction and response through dynamic model consensus. This approach addresses two critical challenges in disaster management by integrating with Digital Twin infrastructures and spatio-temporal remote sensing data. First, handling heterogeneous data distributions across different disaster phases from pre-event to recovery. Second, enabling real-time model adaptation to evolving ground conditions. The method operates through a two-level architecture beginning with multi-model ensemble training, where deep learning models for flood prediction are evaluated via game-theoretic performance attribution using Shapley Values to quantify their context-specific utility. This is followed by a dynamic model selection phase that intelligently matches real-time remote sensing observations against both historical loss patterns and current model competencies, optimizing the ensemble for emerging disaster conditions. Our validation studies demonstrate GTCDL's effectiveness in flood prediction scenarios using Sentinel-1 SAR time series within a river basin Digital Twin environment. The framework achieved 92% AUC in damage forecasting, representing a 6% improvement over static baselines. The full code is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/YousIA/GTCDL/</uri>