ABSTRACT

Water monitoring is an important part of water resource protection. The extraction of water body from multispectral remote-sensing images has been proven to be an efficient and fast way for water monitoring. This paper presents a water body extraction algorithm from multispectral remote-sensing image based on region similarity and boundary information by combining adaptive band selection and over-segmentation. First of all, three bands are adaptively chosen by similarity-based band selection algorithm. Then, the image domain is partitioned into a series of homogeneous sub-regions by over-segmentation incorporating spectral and spatial information. On the sub-regions, the regional similarity is defined with respect to the similarities of texture and spectral features which are extracted using structure analysis method. After that, boundary information is extraction by Canny algorithm, then the water body is extracted by using the Fractal Net Evolution Approach (FNEA) which combines regional similarity and boundary information. The proposed algorithm is used to extract six water bodies with different complex texture backgrounds from multispectral sensors. According to the accuracy evaluation of water body extraction results, the overall accuracy (OA) is higher than 97.9100% and all Kappa coefficients (K) are up to 0.9436. We calculated the relative error (RE) of the area between the reference water body and the water body extracted by the proposed algorithm, the minimum and maximum relative error range is between [0.6180%, 7.7050%]. The experiments show that the proposed algorithm is feasible and effective.

摘要

水资源监测是水资源保护的重要组成部分。从多光谱遥感图像中提取水体已被证明是一种高效，快速的水监测方法。提出了一种基于区域相似度和边界信息的多光谱遥感图像水体提取算法，该算法结合了自适应谱带选择和过度分割。首先，通过基于相似度的频带选择算法自适应地选择三个频带。然后，通过合并光谱和空间信息的过度分割，将图像域划分为一系列均匀的子区域。在子区域上，相对于使用结构分析方法提取的纹理和光谱特征的相似性定义了区域相似性。之后，利用Canny算法提取边界信息，然后利用分形网络演化方法（FNEA）提取水体，该方法结合了区域相似度和边界信息。该算法用于从多光谱传感器中提取具有不同复杂纹理背景的六个水体。根据水体提取结果的准确性评估，总体准确性（OA）高于97.9100％，所有Kappa系数（K）最高为0.9436。我们计算出参考水体与通过提出的算法提取的水体之间的区域的相对误差（RE），最小和最大相对误差范围在[0.6180％，7.7050％]之间。实验表明，该算法是可行和有效的。