The research, conducted by Abdulrhman M. Almoadi of King Abdulaziz City for Science and Technology (KACST) in Riyadh, Saudi Arabia, introduces the Enhanced Normalized Difference Water Index, or ENDWI. It builds on one of the most widely used tools in remote sensing, the Normalized Difference Water Index proposed by S. K. McFeeters in 1996. The classic NDWI exploits a fundamental property of water: it reflects green light strongly but absorbs near-infrared radiation. By combining the green and near-infrared bands in a normalized ratio, NDWI produces high values over water and low values over most land. The trouble is that built-up surfaces, such as concrete roofs and paved roads, also reflect more green than near-infrared, landing in the same spectral neighborhood as water and triggering false positives across urban landscapes.

ENDWI attacks this confusion with a single additional operation: it divides the NDWI value by the green band reflectance. The logic rests on a subtle difference between how water and urban surfaces behave. Both may reflect green light similarly, but water’s strong near-infrared absorption drives its NDWI value far higher than that of rooftops or asphalt, which reflect energy in both bands. When NDWI is divided by the green band, water pixels retain proportionally larger values while non-water urban surfaces are strongly attenuated. The green band effectively acts as a suppressor of urban noise, amplifying the contrast that near-infrared absorption creates. Because the method relies only on the green and near-infrared bands, the same bands used by NDWI, it works with virtually any modern satellite or drone platform, avoiding the shortwave infrared bands that some competing indices require and that are less commonly available on low-cost sensors.

But the elegant formula came with an unexpected mathematical headache. Dividing by the green band compresses ENDWI’s raw values into an extraordinarily narrow range near zero, spanning roughly from minus 0.000412 to plus 0.000150, more than three orders of magnitude tighter than typical spectral indices. That near-zero clustering breaks Otsu’s method, the standard automatic thresholding algorithm that separates water from land by finding the optimal split in a bimodal histogram. To solve this, Almoadi invented a second novel technique called Zero-Preserving Split Normalization, or Z-Split. The method independently rescales the positive and negative portions of the ENDWI distribution to the intervals from zero to one and minus one to zero, while strictly preserving zero as the natural boundary between water and non-water pixels. The result is a cleanly bimodal histogram that Otsu’s method can split reliably. Importantly, Z-Split is not part of the index itself but a preprocessing step, and the author suggests it could benefit any bipolar spectral index suffering from similar value compression.

To test the new approach, the study turned to a dramatic real-world event. On 23 November 2018, heavy rainfall in the upstream catchment of Wadi Al-Lith triggered the partial breach of an earthen retaining dam on Saudi Arabia’s Red Sea coast. A surge of floodwater reached the urban areas of Al-Lith Governorate within roughly four hours, submerging roads and low-lying neighborhoods, with water levels exceeding 1.7 meters in some sectors after further rain on 25 November. Al-Lith’s mix of residential buildings, paved roads, bare soil, and scattered vegetation makes it a textbook example of the spectral confusion that plagues optical flood detection, since water, shadows, and built-up surfaces can appear nearly identical in multispectral imagery.

The validation design was unusually rigorous. Almoadi used Sentinel-2 imagery acquired at 10-meter resolution on 28 November 2018 as the analytical dataset, and validated results against WorldView-4 imagery pan-sharpened to 0.31-meter resolution, acquired within two days of the event. From manually digitized flooded and non-flooded polygons, buffered inward by 10 meters to avoid mixed edge pixels, the study generated 1,262 ground-truth points: 559 flooded and 703 non-flooded. That validation density, concentrated in a compact area of roughly 18 square kilometers, lends considerable statistical weight to the reported accuracies.

The results unfolded in three stages. First, a receiver operating characteristic analysis of eight raw indices, including ENDWI and seven established competitors such as MNDWI, AWEIsh, AWEInsh, WI, LSWI, and SWI, showed the shadow-aware AWEIsh leading with an area under the curve of 0.908, followed by NDWI at 0.671 and ENDWI at 0.636. Second, after Otsu thresholding, ENDWI stole the show among individual indices: it achieved the highest precision at 79.41 percent and the lowest false alarm rate at 10.95 percent, roughly half the false alarm rate of AWEIsh and about one-third that of NDWI. Third, and most strikingly, the study fused the two complementary indices using a pixel-wise maximum operation, taking whichever value, ENDWI or AWEIsh, indicated water more strongly at each pixel. AWEIsh contributed broad separability and shadow handling, while ENDWI suppressed false positives from dark impervious surfaces. The fused map, thresholded with Otsu’s method, achieved 82.65 percent overall accuracy, 94.50 percent precision, an F1-score of 76.73 percent, and a Kappa coefficient of 0.637, with only 21 false positives out of 703 non-flooded validation points, a false alarm rate of 2.99 percent.

Those numbers translate into a 15-fold reduction in false alarms compared with the classic NDWI pipeline, achieved at essentially zero additional computational cost. No deep learning models, no labeled training datasets, no extra sensors, and no multi-sensor fusion are required, just one division operation and a maximum rule that can run in seconds in standard GIS software. That efficiency matters enormously during rapid disaster response, when deep learning approaches, however powerful, demand labeled data and computational resources that are often scarce in the critical first hours after a flood. The method’s reliance on free Sentinel-2 data also makes it attractive for resource-constrained agencies in arid and semi-arid cities, where flash floods are becoming more frequent and severe as climate change intensifies extreme rainfall. Floods affected 2.3 billion people between 1995 and 2015, and between 1980 and 2009 they killed more than half a million people worldwide, making them the deadliest natural disaster type in that period.

The study is candid about its limits. It rests on a single post-event image pair from one flood event, so generalization across seasons, turbidity levels, and vegetation conditions remains unproven, though a preliminary visual check against an independent flood event reported elsewhere suggested ENDWI again outperformed NDWI. Optical methods also require cloud-free skies, and Sentinel-2’s 10-meter resolution can miss narrow urban water features. Future work, the author suggests, could explore adaptive fusion weights, additional bands such as red-edge for vegetation masking, time-series validation across multiple events, and integration with radar data for all-weather capability. Still, the core lesson is a refreshing one for a field increasingly dominated by heavyweight machine learning: sometimes a modest, interpretable change to a 30-year-old formula, applied to free data, can deliver a dramatic practical win. The code, ground-truth points, and the Z-Split Python package have been released openly on GitHub and PyPI, inviting flood mappers everywhere to put the new index to the test.

Subject of Research: An enhanced spectral water index and hybrid fusion method for reducing false alarms in satellite-based urban flood detection

Article Title: Reducing false alarms in urban flood detection: an enhanced NDWI (ENDWI) with Hybrid Max Fusion on Sentinel-2 Data

Article References: Almoadi, A. M. (2026). Reducing false alarms in urban flood detection: an enhanced NDWI (ENDWI) with Hybrid Max Fusion on Sentinel-2 Data. Earth Observation, 1(1), 43-57. https://doi.org/10.5194/eo-1-43-2026

Image Credits: AI Generated

DOI: 10.5194/eo-1-43-2026

Keywords: urban flooding, flood detection, remote sensing, Sentinel-2, NDWI, water index, Otsu thresholding, false alarms, satellite imagery, disaster response, Saudi Arabia, spectral indices

Researchers led by Sk Tahsin Hossain and Tan Yigitcanlar of Queensland University of Technology, working with colleagues at the Chinese Academy of Sciences and Peking University, have unveiled an Artificial Intelligence of Things, or AIoT, enabled urban platform designed to automate flood detection, exposure assessment and impact mapping. The concept is deceptively simple but technically ambitious: instead of relying on slow, manual overlays of flood maps against property datasets, the platform ingests satellite imagery, runs a pretrained deep learning model to delineate inundation, and then automatically cross-references the resulting flood polygons with authoritative data on suburbs, transport infrastructure, buildings and population. The entire pipeline operates within the Esri ArcGIS ecosystem, delivering results through interactive dashboards that refresh as new imagery is processed.

The stakes are considerable. Between 1990 and 2022, more than 4,700 flood events were documented across 168 countries, affecting over 3.2 billion people, killing 218,353 and causing economic losses exceeding 1.3 trillion US dollars. Floods are Australia’s second-deadliest natural hazard after heatwaves, accounting for roughly 20 percent of all hazard-related deaths since 1900, and Queensland is the nation’s most flood-prone state. A statewide analysis by the Queensland Reconstruction Authority found that 60 percent of local councils face high flood risk, while the Climate Council reports that around 70 percent of Queenslanders have experienced at least one flood in the past five years, the highest exposure rate of any Australian state.

What distinguishes the new platform from existing smart-city systems is its insistence on closing the loop between detection and consequence. The authors note that most current flood platforms focus on delineating flood extent and stop there, leaving the critical question of what and who is affected to laborious human analysis. Their architecture is organised into three interdependent layers typical of AIoT frameworks: a sensing layer that can accommodate satellite remote sensing, IoT water-level and rainfall sensors, drone imagery and weather stations; an edge or fog layer for rapid preprocessing and anomaly detection; and a cloud platform layer where deep learning models perform flood extent extraction and object-level mapping before pushing results to decision-support dashboards. For the demonstration, satellite imagery served as the primary sensing source, but the researchers stress that the downstream components are deliberately agnostic to the sensing modality, allowing real-time IoT sensor streams to be plugged in later without redesigning the workflow.

At the heart of the system sits a remarkable piece of open science: the Prithvi-Flood Segmentation model, co-developed by NASA and IBM and distributed through the ArcGIS Living Atlas. Prithvi is a Vision Transformer trained with a self-supervised masked autoencoder strategy and fine-tuned on the Sen1Floods11 benchmark, a collection of 446 labelled image chips drawn from 14 biomes, 357 ecoregions, six continents and 11 major flood events. The model ingests a six-band composite from Harmonised Sentinel-2 or Landsat imagery, ordered as blue, green, red, narrow near-infrared and two shortwave-infrared bands, and classifies every pixel as no water, flood water, or no data and clouds. Its reported performance is striking: an overall accuracy of 97.25 percent, a mean Intersection-over-Union of 88.68 percent, and a flood-water IoU of 80.46 percent. Crucially, when tested on an unseen flood event in Bolivia, the model retained a mean IoU of 86.7 percent, evidence of genuine generalisation across geographic and hydrological conditions.

To demonstrate the platform, the team turned to the March 2022 floods, selecting 62 suburbs across Logan and the Gold Coast, 34 from Logan and the remainder from the Gold Coast, covering both directly inundated areas and adjacent communities at risk. Suburb boundaries were dissolved into a single administrative polygon that acted as a clipping mask for the satellite imagery. The clipped GeoTIFF was then fed to the Prithvi model through the Classify Pixels Using Deep Learning tool, executed on a GPU with test-time augmentation enabled. The classified raster was reclassified, converted to polygons, enriched with attributes, and published as hosted feature layers in ArcGIS Online. In the demonstration setup, equipped with 64 gigabytes of system memory and an NVIDIA A40 GPU, the segmentation step completed in under ten minutes for the entire area of interest.

The real innovation lies in what happens next. Five analytical modules transform raw flood extent into operational intelligence. Module A intersects flood polygons with suburb boundaries and calculates the geodesic flooded area per suburb. Module B classifies every railway station as Already Affected, At Risk or Not Affected, using a 300-metre buffer around the flood extent as the risk threshold, a distance that can be tuned to local conditions. Module C applies the same logic to building footprints, Module D to road segments, flagging potentially impassable routes to support closure and detour decisions, and Module E overlays population data with flood extents to estimate how many residents are exposed. All outputs flow into a five-page dashboard built in ArcGIS Experience Builder, one page per module, giving planners and emergency responders an interactive, continuously updated picture of the disaster.

The researchers are careful about what their system does and does not claim. The platform performs automated exposure assessment and impact mapping, not damage estimation; calculating physical damage, economic losses or casualties would require flood depth, flow velocity, structural vulnerability and asset values that lie beyond the current scope. Nor did the team conduct an independent quantitative validation of the Prithvi model against Queensland-specific reference flood maps, so the case study should be read as a demonstration of the architecture rather than a formal evaluation of model accuracy under local conditions. Other limitations are candidly acknowledged: cloud-free satellite imagery may not be available in the critical hours after a flood, potentially requiring subscription services such as Planet, ICEYE or Maxar, and the event-driven nature of floods makes automated scheduling tricky. The authors suggest coupling the platform to hydrological and meteorological monitoring so that workflows trigger only when rainfall or water-level thresholds are exceeded.

Even with these caveats, the implications are far-reaching. Because the architecture separates sensing, processing and decision-support into modular, interchangeable components, it can be redeployed in any jurisdiction with suitable imagery, administrative boundaries, building footprints, transport networks and population data. The same pipeline could incorporate road extraction models, infrastructure vulnerability assessment or population displacement modelling, and the authors envision evolution towards digital twins, continuous IoT sensor integration and predictive hydrometeorological forecasting, shifting the system from post-event mapping to proactive early warning. Repeated observations of flood exposure could also help governments identify persistently affected locations, evaluate mitigation measures and prioritise investment in resilient infrastructure and nature-based solutions.

As climate change intensifies extreme rainfall, with CSIRO and Bureau of Meteorology projections indicating that the present one-in-100-year flood threshold for Brisbane could be exceeded far more often by mid-century, and modelling suggesting today’s one-percent annual exceedance level could rise by 1.2 to 2.5 metres by 2050, tools that compress the time between satellite pass and actionable intelligence may become indispensable. The Queensland platform offers a glimpse of what disaster response could look like when cities stop reacting to floods with spreadsheets and start answering, automatically and in near-real time, the question that matters most: who and what is in the water’s way.

Subject of Research: An AIoT-enabled urban platform for automated flood detection, exposure assessment and impact mapping using satellite imagery and deep learning

Article Title: AIoT-enabled urban platform for flood detection and impact mapping: towards near-real-time spatial decision support in disaster management

Article References: AIoT-enabled urban platform for flood detection and impact mapping: towards near-real-time spatial decision support in disaster management. (n.d.). https://doi.org/10.1007/s11069-026-08403-3

Image Credits: AI Generated

DOI: 10.1007/s11069-026-08403-3

Keywords: flood detection, AIoT, deep learning, satellite imagery, disaster management, impact mapping, platform urbanism, smart cities, Queensland floods, Prithvi model, geospatial analytics, decision support