Abstract: Synthetic aperture radar (SAR) has the characteristics of all-day and all-weather imaging, wide observation range, and short mapping period, which make it highly advantageous in water extraction. However, existing algorithms for lake extraction are easily affected by the surrounding environment of lakes and noise interference, resulting in low operational efficiency. Therefore, this paper proposes a detection method that combines an improved Gaussian mixture model (GMM) with graph cut model (GCM). First, the two-level Otsu threshold method is used to obtain the initial segmentation map of the lake, and the calculated parameter set is used as the initial parameter of the GMM. The expectation maximum algorithm (EM) is employed to obtain the optimal parameters of the GMM iteratively. The experimental results demonstrate that the more accurate the initial parameters, the clearer the outline of the water body. The introduction of the two-level Otsu algorithm not only greatly reduces the times of iterations of the EM algorithm, but also effectively enhances the running speed of the algorithm in combination with downsampling in preprocessing. In addition, the energy function of the graph cut model enables accurate lake boundaries to be obtained without requiring any post-processing.

Key words: synthetic aperture radar, Gaussian mixture model, graph cut model, two-level Otsu threshold