Threshold Sensitivity and Reference Uncertainty in Sentinel-1 Inundation Mapping: The July 2021 Mahad Flood

Shreyas Khobragade , Kamraan Mulani, Diksha Varekar

Synthetic Aperture Radar (SAR) is widely used to map flood inundation under cloud cover, but backscatter-based mapping is sensitive to processing choices, and quantitative evaluation depends on reference data that are often unavailable. This paper examines both issues for the July 2021 flood at Mahad, Maharashtra, India. Three Sentinel-1A VV/VH scenes from one descending orbit, acquired on 10 July, 22 July and 3 August 2021, were used to derive backscatter change in Google Earth Engine. Candidate masks were generated with a fixed −2 dB threshold and with Otsu-derived thresholds (about −2.39 dB), after provisional exclusion of long-term water. The two masks covered approximately 10.8819 and 9.4349 km² with a mutual intersection over union (IoU) of 0.8670, which illustrates threshold sensitivity rather than accuracy. The NRSC Bhuvan portal displayed an inundation layer for 22 July 2021, but the original event-specific GIS data could not be obtained through the public access routes investigated, and a request sent to NRSC on 21 October 2026 had not been answered at the time of writing. As a documented workaround, the portal display was captured as a screenshot, georeferenced in QGIS using 11 ground control points, and converted to a provisional binary reference mask by RGB thresholding. On a common 662 × 649 grid, this reference contained 27,317 positive pixels (6.5590 km²) and the aligned Sentinel-1 shared candidate mask 38,597 (9.2675 km²); they shared 17,749 pixels (4.2617 km²), giving an IoU of 36.85%. This value measures spatial agreement with a screenshot-derived reference. It is neither the original NRSC product, nor independently verified ground truth, nor an estimate of flood-detection accuracy. The study is an exploratory analysis, not a validated flood-extent product or a risk assessment.