Continuous Flood Nowcasting in South Asia: A Multi-Sensor Ensemble Remote Sensing Framework for Flood Extent
Usman Nazir, Disha Gomathinayagam, Muhammad Kamran, Sara Khalid · 2026
Read Continuous Flood Nowcasting in South Asia: A Multi-Sensor Ensemble Remote Sensing Framework for Flood Extent by Usman Nazir, Disha Gomathinayagam, Muhammad Kamran, Sara Khalid (2026) 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-2605-10950.txt.
Introduction
Usman Nazir, Disha Gomathinayagam, Muhammad Kamran, and Sara Khalid build a continuous flood-nowcasting system for Pakistan, where the 2025 monsoon season ran from June to December with cascading damage. Existing flood products give episode snapshots, but rarely continuous, near-real-time inundation maps within the country.
Their ensemble fuses Sentinel-1 radar with Landsat-Sentinel, MODIS, and VIIRS imagery on a harmonized grid, preferring high-resolution sensors and falling back to coarser ones to keep daily coverage. A day-by-day track of the late-August super-flood shows the footprint evolving in near real time, checked against discharge and rainfall data. Disaster responders get maps that keep up with the water, not just summaries after it.
Catalog record
| CREATOR | Usman Nazir, Disha Gomathinayagam, Muhammad Kamran, Sara Khalid |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 31370 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.10950
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Continuous Flood Nowcasting in South Asia: A Multi-Sensor Ensemble Remote Sensing Framework for Flood Extent Usman Nazir, Disha Gomathinayagam, Muhammad Kamran, Sara Khalid Abstract Pakistan experienced an unusually severe flood season between June and December 2025, with cascading impacts on population, infrastructure, and agriculture. Existing operational flood products (e.g., UNOSAT) provide valuable episode-level snapshots but rarely deliver spatially and temporally continuous inundation maps at near-real-time latency within the country. We present a multi-sensor, ensemble-based remote-sensing framework for continuous flood nowcasting in Pakistan that integrates Sentinel-1 SAR, Harmonized Landsat–Sentinel (HLS L30 and S30), MODIS, and VIIRS observations on a harmonized grid in Google Earth Engine. The framework employs a tiered nowcasting ensemble that prioritizes higher-resolution sensors (Sentinel-1 and HLS) and falls back to MODIS and VIIRS when necessary, preserving daily continuity of flood extent at each sensor’s native resolution. Applied to the 2025 monsoon period, the system generates near-real-time, spatially consistent inundation maps across Pakistan. As a nowcasting case study, we track the super-flood of 26 August–7 September 2025 day by day, demonstrating the framework’s ability to capture the evolving flood footprint in near real time and extend beyond the temporal limits of episodic mapping products. Validation against GloFAS discharge anomalies and precipitation datasets (CHIRPS v3.0, MSWEP) shows strong agreement with observed hydrometeorological conditions. By integrating nowcast outputs with exposure layers (WorldPop, ESA WorldCover, Giga-HOTOSM), the framework enables rapid estimation of affected populations, cropland, and critical infrastructure, supporting timely disaster response and resilience planning in South Asia. 1 Introduction Floods are among the most destructive climate-related hazards in South Asia, driven by a combination of monsoonal precipitation, glacial melt, tropical cyclones, and rapid urbanization in exposed flood plains. Timely and spatially complete flood extent information is essential for humanitarian response, risk assessment, and climate adaptation planning. However, operational products such as UNOSAT flood maps are typically produced on an episode basis for specific events and specific countries, leaving substantial gaps in both space and time. In practice, responders need a nowcasting capability: updated flood maps with latency measured in days, not weeks, and continuous coverage between episodes rather than single post-event snapshots. We develop and apply a cloud-based, multi-sensor ensemble framework for continuous flood nowcasting across Pakistan during the 2025 flood season, and combine the resulting inundation maps with population, land cover, and infrastructure data to quantify exposure. Because flood seasonality differs across the region—five of the six countries experience the bulk of their flooding during the southwest monsoon, while Sri Lanka’s 2025 flooding was concentrated in the northeast-monsoon and cyclone-driven late-season window—we adopt country-specific observation windows : JJASO (June–October) 2025 for Pakistan, India, Nepal, Bangladesh, and Bhutan, and November–December 2025 for Sri Lanka. The combined reporting window is therefore June–December 2025. To illustrate the nowcasting behaviour of the framework, we present an in-depth case study of the Pakistan super-flood of 26 August–7 September 2025 (Section 5.1 ). Contributions. • A reproducible Google Earth Engine pipeline that harmonizes Sentinel-1 SAR, HLS Landsat-30, HLS Sentinel-2, MODIS, and VIIRS observations and operates a two-mode ensemble: (a) a tiered nowcasting ensemble that preserves high resolution when Tier-1 sensors are available and uses MODIS/VIIRS only as a daily-coverage fallback, and (b) a retrospective 3-of-5 majority-vote aggregation on a 30 m grid for seasonal totals (Section 3 ). • A country-specific observation-window strategy (JJASO for five countries, November–December for Sri Lanka) that aligns the nowcasting cadence with regional flood seasonality. • A nowcasting case study of the Pakistan 26 August–7 September 2025 event (Section 5.1 , Figure 3 and Figure 4 ), showing that our framework reproduces the UNOSAT footprint in near-real-time and extends it across the full monsoon window. • A 242-event hazard inventory assembled from ReliefWeb, multi-source reports, and field expert validation, classifying riverine, flash, urban, coastal, cyclone, monsoon, landslide, and compound events (Section 4 ). 2 Related Work Operational flood mapping. UNOSAT and the Copernicus Emergency Management Service (EMS) Rapid Mapping products provide event-triggered flood extents using SAR and optical imagery, with turnaround in hours to days. They are widely used for humanitarian response but are activated on a per-event basis, leaving temporal and spatial gaps across a full monsoon season and unequal country coverage ( 14 ; 5 ) . SAR-based flood detection. Sentinel-1 backscatter thresholding is a mature approach for open-water detection under cloud cover ( 10 ; 1 ) . Limitations include backscatter ambiguity in urban areas, flooded vegetation, wet soils, and dry sand. Recent work has explored change-detection and data-fusion strategies to mitigate these effects. Optical water indices. NDWI ( 11 ) , MNDWI ( 15 ) , and AWEI ( 6 ) are standard spectral water indicators used with Landsat and Sentinel-2 observations; Harmonized Landsat–Sentinel (HLS) products ( 4 ) enable dense time series on a common grid. Multi-sensor and learned approaches. Ensemble fusion of SAR and optical inputs has been shown to reduce sensor-specific bias. Learned segmentation models trained on benchmarks such as Sen1Floods11 ( 3 ) and the ETCI-2021 challenge ( 12 ) represent a complementary direction; our framework is deliberately ensemble-and-threshold based to preserve interpretability, run directly in Google Earth Engine, and avoid labelled-data requirements in countries where ground truth is sparse. Hydrological and meteorological reference data. We validate against GloFAS discharge reanalysis ( 9 ) and the CHIRPS ( 7 ) and MSWEP ( 2 ) precipitation datasets, which are standard choices for regional flood and extreme-rainfall analysis. 3 Proposed Methodology 3.1 Study Design and Temporal Framework Flood extent is derived using a multi-sensor remote-sensing framework integrating synthetic aperture radar (SAR) and optical satellite observations. A pre-flood (baseline) period and a flood-impact period are defined using rolling temporal windows. Satellite observations within each period are filtered and aggregated using median compositing to reduce noise, residual cloud contamination, and short-term variability. All analyses are conducted in Google Earth Engine. Nowcasting windows. To support operational nowcasting, the pre-flood baseline and the flood-impact window are updated on a rolling basis as new Sentinel-1, HLS, MODIS, and VIIRS acquisitions arrive. Concretely, for a given nowcast timestamp t we define a pre-flood composite over [t-\Delta_{\text{pre}},t-\Delta_{\text{lead}}] and a flood composite over [t-\Delta_{\text{flood}},t] , with defaults \Delta_{\text{pre}}=60 days, \Delta_{\text{lead}}=14 days, and \Delta_{\text{flood}}=7 days. These windows are applied independently for each sensor and are shifted forward at each update; this produces a near-real-time flood extent product with latency governed by satellite revisit, downlink, and Google Earth Engine ingest (on the order of 1–3 days for Sentinel-1 and HLS over our region). Section 5.1 illustrates this rolling nowcast behaviour on the Pakistan 26 August–7 September 2025 event. Study region. The study region is defined using FAO GAUL administrative boundaries or custom subregional geometries. All datasets share a common geographic coordinate reference system (EPSG:4326). Resampling to a 30 m common grid is applied for the retrospective Mode B ensemble (Section 3.3 ); for the Mode A tiered nowcast, per-sensor binary masks are retained at their native resolutions and combined through the tiered preference rule described in Section 3.3 . Figure 1 : Proposed multi-source, two-mode ensemble methodology for flood extent nowcasting and retrospective mapping. The workflow ingests multi-sensor satellite data (Sentinel-1 SAR, HLS Landsat-30, HLS Sentinel-2, MODIS, and VIIRS) over a defined study region and time period, organised into Tier 1 (S1, L30, S30)