Decision-Focused Sensor Network Placement for Rapid Disaster Response

Timely and reliable decision-making is critical for disaster emergency response, yet it is fundamentally constrained by limited situational awareness due to sparse and imperfect sensing under budget and data accessibility constraints. In current practice, flood management systems follow a two-stage pipeline that separately reconstructs and predicts flood states from sparse sensor observations, and then optimizes downstream response decisions, such as evacuation planning and resource allocation. However, existing approaches typically adopt decision-agnostic strategies for both sensor placement (e.g., maximizing information gain) and model training (e.g., minimizing average prediction error), neglecting the inherent mismatch between sensing effectiveness or predictive accuracy and decision quality. As a result, systems with similar sensing performance or forecasting errors can yield substantially different decisions. To address this limitation, we propose a decision-focused learning (DFL) framework for sensor-aware flood monitoring and response optimization. First, we formulate sensor placement and flood forecasting as a unified decision-informed learning problem, where sensing, prediction, and optimization are tightly coupled. Second, we train the reconstruction and forecasting model to directly minimize downstream decision regret, enabling end-to-end alignment between prediction and optimization objectives. Third, to handle the heterogeneous optimization structures in the pipeline, namely discrete combinatorial sensor selection under hard budget constraints and structured downstream decision-making, we develop a hybrid DFL framework that integrates perturbation-based learning via Implicit Maximum Likelihood Estimation (I-MLE) for differentiable sensor selection with surrogate-based decision-focused optimization (SPO+) for decision-aligned model training. Extensive experiments on real-world hurricane and flood scenarios demonstrate that the proposed framework consistently improves downstream decision quality, achieving lower regret compared to prediction-focused baselines. These results highlight the importance of decision-aware sensing and provide a principled approach for integrating sensor placement, forecasting, and optimization in disaster response systems.