Where to Watch the Water: Multi-Sensor Network Design Optimization for Inland Flood Detection

Basit Akinwumi Akinade, Amobichukwu Chukwudi Amanambu, Lisa Davis

Inland flood detection is often constrained less by sensor availability than by where sensors are placed along branching
river networks, especially in ungauged headwaters where floods often initiate. We present a three-phase, decisionfocused
framework for designing basin-by-basin multi-sensor flood detection networks that coordinate water-level,
discharge, and camera sensors while explicitly balancing risk-weighted detection, false-alarmpenalties, and deployment
cost. Using greedy submodular optimization across 65 HUC10 basins in the Southern Appalachians, we generated
four operational scenarios spanning coverage-maximizing to resource-constrained deployments, with 491 to 216
potential sensor locations, respectively, with observed diminishing returns. Across scenarios, optimized networks
consistently outperform common allocation baselines (random and uniform-grid placement), with improvements over
random placement rising from 36.8% under the maximum-coverage design to 77.3% under the resource-constrained
design, and remaining about 59.5% under the intermediate configurations. We then validate hydrologic relevance
using 44 years (1979–2023) of National Water Model Retrospective v3.0 streamflow, showing that selected sites
are systematically located on higher-flow river segments relative to a spatially independent reference set. Finally, we
quantify complementarity with 69 existing USGS gages, demonstrating that optimized networks act as modular add-ons
that expand risk-weighted coverage from 44.3% (USGS-only) to as high as 96.7% when combined, while supporting a
cascade warning architecture with median upstream lead times of ∼2 hours. The results provide a scalable framework
for expanding flood monitoring to maximize early-warning value per sensor deployed.