Decision-focused Sensing and Forecasting for Adaptive and Rapid Flood Response: An Implicit Learning Approach Conference: The 12th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation; November 19–21, 2025; Golden, CO, USA The 12th ACM International Conf

Qian Sun, Graham Hults, Susu Xu · 2025

Read Decision-focused Sensing and Forecasting for Adaptive and Rapid Flood Response: An Implicit Learning Approach Conference: The 12th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation; November 19–21, 2025; Golden, CO, USA The 12th ACM International Conf by Qian Sun, Graham Hults, Susu Xu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-16015.txt.

Introduction

Qian Sun, Graham Hults, and Susu Xu design flood sensing and forecasting around the decisions they must support. Fast, reliable choices save lives in flood response, yet budgets and access limits leave awareness thin. Standard practice places ground sensors to gain the most information overall and trains forecast models to minimise average error, ignoring that equal scores can still lead to very different decisions.

Their end-to-end framework picks sensor sites and trains a spatio-temporal forecast model to cut downstream decision regret directly. A contextual scoring network, a differentiable sensor selector under a hard budget, a flood reconstruction and forecasting model, and a differentiable decision layer work as one pipeline. Implicit maximum likelihood estimation lets gradients flow through discrete sensor choices, and probabilistic decision heads approximate varied constrained response tasks in differentiable form.

Catalog record

CREATORQian Sun, Graham Hults, Susu Xu
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE55 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-16015

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Decision-focused Sensing and Forecasting for Adaptive and Rapid Flood Response: An Implicit Learning Approach Conference: The 12th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation; November 19–21, 2025; Golden, CO, USA The 12th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation (BUILDSYS ’25), November 19–21, 2025, Golden, CO, USA DOI: 10.1145/3736425.3770102 ISBN: 979-8-4007-1945-5/2025/11 CCS: Computing methodologies Modeling and simulation CCS: Hardware Sensor applications and deployments CCS: Applied computing Forecasting CCS: Applied computing Decision analysis Qian Sun, Graham Hults, Susu Xu Abstract . Timely and reliable decision-making is vital for flood emergency response, yet it remains severely hindered by limited and imprecise situational awareness due to various budget and data accessibility constraints. Traditional flood management systems often rely on in-situ sensors to calibrate remote sensing-based large-scale flood depth forecasting models, and further take flood depth estimates to optimize flood response decisions. However, these approaches often take fixed, decision task-agnostic strategies to decide where to put in-situ sensors (e.g., maximize overall information gain) and train flood forecasting models (e.g., minimize average forecasting errors), but overlook that systems with the same sensing gain and average forecasting errors may lead to distinct decisions. To address this, we introduce a novel decision-focused framework that strategically selects locations for in-situ sensor placement and optimize spatio-temporal flood forecasting models to optimize downstream flood response decision regrets. Our end-to-end pipeline integrates four components: a contextual scoring network, a differentiable sensor selection module under hard budget constraints, a spatio-temporal flood reconstruction and forecasting model, and a differentiable decision layer tailored to task-specific objectives. Central to our approach is the incorporation of Implicit Maximum Likelihood Estimation (I-MLE) to enable gradient-based learning over discrete sensor configurations, and probabilistic decision heads to enable differentiable approximation to various constrained disaster response tasks. Rather than minimizing proxy prediction loss, our method directly optimizes final task performance metrics such as evacuation and relocation cost. Extensive experiments on real-world flood scenarios demonstrate that our method substantially outperforms baselines in both state prediction accuracy and decision quality, offering a principled and effective solution for sensor-aware disaster management. 1. Introduction Rapid and effective decision making in response to natural hazards (e.g. floods, hurricanes, wildfires, etc.), including massive evacuation, resource allocation, and facility relocation, is critical to reducing human and economic loss ( xu2022seismic ; wang2024scalable ; zhou2018emergency ) . To enable near-real-time decision-making, traditional paradigms often (1) first employ sensing systems to collect spatio-temporal data of disaster zones, (2) then use physical, statistical, or machine learning models to estimate or forecast spatio-temporal damage distribution, and finally (3) optimize various response decisions and dynamically update decisions as natural hazard damage evolves ( wang2024near ; chen2024soscheduler ; li2023disasternet ; huang2021systematic ) . For example, in flood management, decision-makers often rely on multi-modal sensing data (e.g., precipitation, soil moisture, elevation, etc.) and machine learning models to estimate spatio-temporal distribution of flood depth, and then using these flood depth as decision parameters to optimize dynamic evacuation and facility relocation in hours to days before flooding to reduce loss. In particular, due to the large-scale and fast-changing nature of natural hazard damage, recent disaster sensing paradigms often involve multiple modalities ranging from remote sensing techniques and in-situ sensors deployed at key locations. Remote sensing data cover large-scale disaster zones, but often suffer from limited spatial resolution, infrequent updates, and poor signal quality ( al2013major ; amitrano2024flood ) , making it difficult to accurately capture local flood dynamics. As complementary information sources, in situ sensors provide accurate and real-time point observations ( munawar2022remote ; tao2024review ) , such as the stage and discharge of water, but often have limited spatial coverage due to high deployment and maintenance costs. In real-world practice, people often fuse large-scale, coarse-grained remote sensing data with sparse but accurate fine-grained in-situ measurements to jointly estimate and forecast the spatio-temporal distribution of flood depth using machine learning models ( liu2025comprehensive ; bukhari2024enhancing ) , where in-situ point observations often play the role of calibrating remote sensing-based estimations. Therefore, the effectiveness of final response decisions are jointly determined by the quality of sensing measurements, forecasting models, and decision optimization algorithms. To enable effective flood response systems , people often need to jointly answer three key questions: (1) where to place in-situ sensors? (considering satellite imagery often has fixed characteristics) (2) how to adapt hazard/damage forecasting models with multi-modal sensing data? (3) how to optimize response decisions with forecasting results? , by simultaneously considering their coupled impacts on final response decision quality. However, existing flood management frameworks often isolate sensor deployment optimization and forecasting model optimization from actual disaster decision-making problems, resulting in sub-optimal outcomes. In fact, disaster responses involve a diverse spectrum of decision tasks, including assignment, matching, routing, and control, each needing tailored sensing strategies and forecasting models. Current approaches for sensing network optimization mostly prioritize maximizing average information gain. Similarly, many data-driven forecasting models focus on minimizing average forecasting errors on sampling datasets ( li2023normalizing ; li2025rapid ; li2025scalable ) , but these objectives do not necessarily improve disaster response decisions ( saunders2025data ; gerolimos2025autonomous ) . In real-world highly uncertain and resource-constrained scenarios, systems with the same sensing coverage and learning-based prediction accuracy may lead to distinct decisions and decision optimality. For example, achieving lower forecasting errors around critical infrastructures can significantly enhance resource allocation, even if overall average performance is unchanged. Traditional sensing objectives (e.g., aggregated information gain) and forecasting objectives (e.g., average training loss) may not coincide with optimal decision regrets, underscoring the need for a decision-focused, end-to-end framework to boost flood resilience. To bridge this significant research gap, we introduce an end-to-end, implicit decision-driven framework that jointly learns to place sensors with limited budget and accessibility constraints, optimize flood estimation models, and make various flood response decisions to improve flood management effectiveness and efficiency. Specifically, our approach consists of four interconnected, decision problem-informed components: (1) a location scoring network that assigns importance scores to candidate locations based on contextual features from satellite data; (2) a sensor sampling module that samples a subset of locations for deployment under hard budget/accessibility constraints; (3) a spatio-temporal reconstruction and forecasting module that reconstructs and predicts spatio-temporal flood depth across the entire watershed based on coarse-grained remote sensing measurements of environmental factors (gridded precipitation), sparse in-situ point observations (water depth and velocity), and geospatial features (digital elevation model, satellite-derived Manning’s n surface roughness values, and satellite-derived distance to nearest waterbody); and (4) a differentiable decision head that takes the predicted flood depth as input and outputs actionable plans tailored to specific flood response tasks. These four components are jointly trained and optimized to minimize the final decision regrets. Key challenges in this decision-focused end-to-end learning framework arise from the inherently discrete and combinatorial nature of both sensor placement optimization and disaster response decision-making. Specifically, these tasks form combinatorial optimization problems characterized by cardinality constraints and non-differentiable discrete choices, thereby precluding direct application of gradient-based methods. Secondly, optimal sensor subset selection from extensive candidate locations, as well as operational decisions like evacuation planning or strategic facility relocation, often involve exponentially large discrete search spaces due to the large-scale nature of natural hazard damage and decision-making. Moreover, joint optimization of these two problems with intermediate flood forecasting

Read the complete text