Han Cao

Unsupervised flood detection in large areas using Synthetic Aperture Radar (SAR) data always faces the challenge of automatic thresholding, because the histograms of large-scale images are unimodal, which thus makes it difficult to determine the threshold. In this paper, an iteratively multi-scale chessboard segmentation-based tiles selection method is introduced. This method includes a robust search procedure for tiles which obey bimodal Gaussian distribution, and a non-parametric histogram-based thresholding algorithm for thresholds identifying water areas. Then, the thresholds are integrated into the region-growing algorithm to obtain a consistent flood map. In addition, a classification refinement technique using multiresolution segmentation is proposed to address the omission in a heterogeneous flood area caused by water surface roughening due to weather factors (e.g., wind or rain). Experiments on the flooded area of Jialing River on July 2018 using Sentinel-1 images show a high classification accuracy of 99.05% through the validation of Landsat-8 data, indicating the validity of the proposed method.

To obtain accurate information in a timely manner on built-up areas (BAs) is essential for urban planning and natural hazard (e.g., earthquakes) response strategies. In this paper, a new method for BAs extraction using the Sentinel-1 SAR is proposed, which includes two steps: (1) Candidate BAs are first selected as seeds from images that show high backscattering and obvious textural patterns, as characterized by image intensity, Getis-Ord index, and the variogram texture features; (2) region growing is iteratively implemented from these seed pixels to extract the BAs. Sentinel-1 data, with 5 × 20 m² resolution, are selected over eight cities with various environmental settings around China, to validate the robustness of the proposed method. The results show that the proposed method achieves higher detection accuracy and fewer commission errors compared with the intensity-based region growing and thresholding methods. An averaged accuracy of 96.5% in validation points of eight cities was achieved, which outperforms the GlobCover urban product in both urban and rural area, while fewer commission errors were achieved compared to Landsat data-based methods. Moreover, two polarizations (VV/VH) and the averaged channel are compared for BAs extraction in areas with various environments. It turns out that improved results can be achieved using the averaged image of two polarizations in north China, while the VV image is better suited for BAs extraction in south. These findings indicate that operational BAs mapping over China, and even globally, is possible, since the Sentinel-1 data can provide images with global coverage.