基于多特征融合的反向传播神经网络高分影像分类与变化检测

为了降低基于高分影像的土地利用分类后的错分和漏分的可能性,提高分类以及变化检测精度,本文以广西桂林市临桂区为研究区,采用WorldView-2号以及高景一号高分影像,基于多层前馈(back propagation, BP)神经网络方法融合遥感影像的纹理、光谱、植被指数以及水体指数特征,制定出4种特征数据集融合方案,实现对植被覆盖率较大地区的地物识别与分类；然后选取最优分类结果,进行桂林市临桂区2017与2020年土地利用变化检测。不同方案的对比结果表明,融合纹理、光谱、植被指数以及水体指数特征的第四种方案可以得到较为有效的分类以及变化检测结果,分类的总体精度为92.92%,Kappa系数为0.9028,保持了较高正确率。

High-Resolution Remote Sensing Image Classification and Change Detection Based on BP Neural Network with Multi-Feature Fusion

In order to reduce the possibility of land use misclassification based on high-resolution remote sensing image, and improve classification and change detection accuracy, Taking Lingui District, Guilin City, Guangxi as the research area, using WorldView-2 and SuperView-1 high resolution remote sensing images, integrating texture features, spectral features, vegetation index and water index features, and designing four types of feature fusion in the scheme, the BP neural network method is used to realize the recognition and classification of features in areas with large vegetation coverage; then select a best classification result to detect land use changes in Lingui District, Guilin City in 2017 and 2020. The comparison results of different schemes show that the scheme 4, which integrates texture features, spectral features, vegetation index and water index features, has more effective classification and change detection results can be obtained. The overall accuracy of classification is 92.92%, and the Kappa coefficient is 0.9028, the correct rate is better maintained.