基于机器学习的多特征融合高分辨率遥感影像土地利用分类研究
Research on Land Use Classification of Multi-Feature Fusion High-Resolution Remote Sensing Images Based on Machine Learning

为了提高土地利用分类精度，本文以高分二号遥感影像作为基础实验数据，融合影像光谱信息、归一化植被指数(NDVI)和纹理信息形成多特征融合影像，分别采用神经网络分类方法和支持向量机分类方法对高分辨率遥感影像进行土地利用分类研究，并对两种分类方法结果进行分类精度对比。研究结果发现：1) 多特征融合影像分类精度优于单独使用研究区遥感影像波段光谱信息进行分类取得的精度，很大程度上提高了土地利用分类准确度。2) 与神经网络分类方法相比，基于多特征融合的支持向量机分类法分类斑块碎化程度较小，图斑完整性较好，地物错分漏分现象较少，并且从总体精度和Kappa系数来看，支持向量机分类法优于神经网络分类，且基于多特征融合影像的SVM分类总体精度达到了93.98%，Kappa系数为0.8981。因此基于多特征融合影像的SVM分类能够有效提高土地利用分类精度，可为土地利用监测和土地整治提供有效的数据和技术支持。
In order to improve the accuracy of land use classification, this paper uses Gaofen-2 remote sensing images as the basic experimental data, and fuses image spectral information, normalized vegetation index (NDVI) and texture information to form multi-feature fusion images, using neural network classification methods respectively. The land use classification of high-resolution remote sensing images is studied with the support vector machine classification method, and the classification ac-curacy of the results of the two classification methods is compared. The research results show that: 1) The classification accuracy of multi-feature fusion images is better than that obtained by using the spectral information of remote sensing image bands in the study area alone, which greatly im-proves the accuracy of land use classification. 2) Compared with the neural network classification method, the support vector machine classification method based on multi-feature fusion has less fragmentation degree, better patch integrity, less misclassification and omission of ground objects, and from the overall. In terms of accuracy and Kappa coefficient, support vector machine classification is better than neural network classification, and the overall accuracy of SVM classification based on multi-feature fusion images reaches 93.98%, and the Kappa coefficient is 0.8981. Therefore, SVM classification based on multi-feature fusion images can effectively improve the accuracy of land use classification, and can provide effective data and technical support for land use monitoring and land remediation.