Adaptive Threshold Algorithms for Real-Time Flood Detection Using IoT Sensors

Abstract

Real-time flood detection demands algorithms that react quickly to hydrologic change while remaining robust to sensor noise, seasonal drift, and connectivity constraints typical of Internet-of-Things (IoT) deployments. Classical static thresholds (e.g., a fixed water-level cutoff) are simple but brittle: they generate false alarms during monsoon build-up and miss fast-rising flash floods when the baseline regime shifts. This manuscript proposes and evaluates an adaptive, multi-criteria thresholding framework that runs on resource-constrained edge nodes and scales to catchment-wide networks. The core idea is to couple Exponentially Weighted Quantiles (EWQ) for dynamic baselines with robust dispersion measures (MAD), rate-of-rise checks, and change-point logic (CUSUM/Page-Hinkley) and then fuse them into a single Risk Index with hysteresis and upstream context. We describe an implementable algorithm using O(1) memory updates and percentile tracking via the P² algorithm, suitable for LoRaWAN/NB-IoT sensors.

A simulation study with synthetic hyetographs and a unit-hydrograph routing model across 20 virtual stations compares the proposed method to static thresholds, moving-average dynamic thresholds, and standalone CUSUM. Results show a median detection latency reduction of 38–55% versus baselines, a false alarm rate below 0.2/day in noisy conditions, and improved F1-scores (0.89 vs. 0.71–0.83). We also quantify energy and bandwidth savings from edge filtering and event-driven reporting. The paper concludes with deployment considerations, limitations (e.g., extreme outliers, sensor drift beyond calibration), and practical guidance for tuning in monsoon-dominated basins.