A novel change homogenizing flood index and an expedited locally tuned Markov random field model for unsupervised rural floodwater detection in sentinel-1 data: An application to flood impact assessment

Due to their all-weather, all-day data ingestion capability, interest in synthetic aperture radar (SAR) remote sensing satellites is growing for rural flood mapping in contemporary times. Current SAR flood detection change indices cannot correct for dependency on reference backscatter values when computing flood changes. To address this fundamental limitation, the present study proposes an unsupervised rural flood detection method that produces flood extent maps in two main steps. In the first step, a novel change homogenizing flood index (CHFI) was proposed to reduce the heterogeneity of flood-change patterns by highly mitigating the aforesaid dependency. In the second step, an expedited locally tuned Markov Random Field (ELTMRF) model was developed to partition the CHFI into flood and non-flood classes within a reasonable time while maintaining boundary accuracy. Finally, to evaluate flood impacts on different LULC classes, a land use/land cover (LULC) map for the flood-affected region was first generated using the random forest classifier and then intersected with the ELTMRF-based flood extent map. In this research, Sentinel-1 and −2 data for eight flood events were gathered to conduct experiments in two stages: (1) the evaluation of CHFI against existing change indices (CIs) and (2) the comparison of ELTMRF to state-of-the-art (SOTA) unsupervised flood detection approaches. According to the experimental results, the proposed CHFI enhanced the homogeneity of the flood-related changes while establishing a satisfactory contrast between non-flood and flood classes. Thereby, by yielding an average F-score value of around 88%, the proposed CI exhibited the most generalizable, superior performance compared to the conventional CIs. The ELTMRF model, preserving the boundaries of flood objects, also improved the prior MRF-based models by approximately 5.69% and 41.6% in accuracy and runtime, respectively. Furthermore, the proposed approach, which combines CHFI and ELTMRF, reached an average IoU of 84.98%, demonstrating a remarkable superiority of nearly 12% and 9% over unsupervised and supervised flood mapping methods, respectively. Regarding the impacts on the land use classes, the analyses revealed that farmlands and roads were the two classes most extensively affected by flooding in all cases. Accordingly, the proposed unsupervised rural flood detection approach, which combines CHFI and ELTMRF, provides an efficient framework for detecting flooded areas in bi-temporal SAR data, achieving high accuracy while requiring reasonable computational time. The datasets and the module of the proposed flood mapping method will be available on https://github.com/AminMohsenifar/CHFI-ELTMRF.