Enhanced large-scale flood mapping using data-efficient unsupervised framework based on morphological active contour model and single synthetic aperture radar image.

Floods are critical hydrological extremes that cause significant environmental damage. Remote sensing data, specifically Synthetic Aperture Radar (SAR), derived flood maps are crucial for detecting and quantifying this damage, enabling effective flood management and damage assessment. However, majority of SAR-based flood mapping frameworks are supervised and often face the problem of data generalisation and data labelling, which presents a challenge for model transferability and rapid mapping. To address this challenge, an unsupervised computer vision-based framework built upon the Morphological Chan-Vese Active contour model (Morph CV ACM) and unitemporal SAR image is proposed for flood extent mapping. In this work, sensitivity analysis of the model parameters is performed to check its applicability for two commonly found flooding patterns (clustered and scattered) in SAR images. Furthermore, a localised version of Morph CV ACM is proposed, which adaptively adjusts the model parameters according to the specific characteristics of flooding patterns, based on an empirically developed formula. The proposed framework is tested to map floods that occurred in North India in 2023 across the flood plains of the Yamuna River. The results were validated against flood reference masks generated by PlanetScope optical images for six different Areas of Interest (AOIs), representing varied land covers and flooding patterns. The novel framework accurately identified flood extents with a high F1 score of 0.935. Flood extents from the proposed framework were also compared with the Otsu segmentation, a widely established unsupervised method, and results indicated a major improvement in detecting flood extents with our framework. The improvements in the performance were attributed to the inherent property of Morph CV ACM to use region-based information to govern the energy equation of the model, leading to accurate flood boundary detection, while its use of morphological metrics enhances resilience to the speckle effect in SAR images. Additionally, the generated flood extents were overlaid on the 10 m resolution land cover map for performance assessment across different land covers. The extents generated from our framework provide enhanced flood mapping accuracy in built-up and agricultural areas, where precise mapping using SAR data is challenging yet crucial for damage assessment. The framework's automation and minimal data requirements make it a valuable tool for near-real-time, large-scale flood mapping, with significant potential to enhance damage assessment and guide effective flood management strategies.