ABSTRACT

Surface water, as the most important land resource, profoundly affects the balance of the ecosystem and the development of social economy. The rapid development of remote-sensing technology provides the possibility of dynamic monitoring and real-time automatic extraction of surface water. Based on the index of ZhuHai-1 Orbita-hyperspectral (OHS) pixels, this paper proposes a supervised learning method for semi-automatic surface water extraction. The method first uses a threshold method to segment the cosine distance between pixel spectra, realizes automatic acquisition of training set labels required for supervised classification and effectively improves work efficiency. We then combine the extraction results of the support vector machine (SVM), which only considers spectral information, and the variant full convolutional neural network (VFCN), which only considers spatial information, to obtain more accurate surface water information. In this paper, the analysis of typical waterbodies in four regions showed that the performance of the surface water extraction for all scenes when using this method in this paper is better than that when using the water index (WI), VFCN or SVM. The overall accuracy of this method is above 98.822%, and the kappa coefficients are above 0.9019.

摘要

地表水作为最重要的土地资源，深刻影响着生态系统的平衡和社会经济的发展。遥感技术的快速发展为地表水的动态监测和实时自动提取提供了可能。本文基于珠海一号轨道高光谱（OHS）像素指标，提出一种半自动地表水提取的监督学习方法。该方法首先采用阈值法对像素谱之间的余弦距离进行分割，实现了监督分类所需训练集标签的自动获取，有效提高了工作效率。然后我们结合只考虑光谱信息的支持向量机（SVM）和变体全卷积神经网络（VFCN）的提取结果，只考虑空间信息，获取更准确的地表水信息。本文通过对四个区域典型水体的分析表明，使用本文方法对所有场景的地表水提取性能优于使用水指数（WI）、VFCN或SVM时。该方法总体准确率在98.822%以上，kappa系数在0.9019以上。