A Land-cover Classification Method of High-resolution Remote Sensing Imagery Based on Convolution Neural Network

With the development of space satellites, a large number of high-resolution remote sensing images have been produced, so the analysis and application of high-resolution remote sensing images are very important. Recently deep learning provides a new method to increase the accuracy of land-cover classification. This study aims to propose a classification framework based on convolutional neural network (CNN) to carry out remote sensing scene classification. After remote sensing images are trained by CNN, a model which can extract complex characteristic from the image for classification is created. In this paper, GaoFen-2(GF-2) satellite data is used as data sources and Jilin province of China is selected as the study area. Firstly, the preprocessed images are made into a GF-2 satellite data sets. Secondly, CaffeNet is used to train the data sets through Caffe platform and the classification result is obtained. The CNN overall accuracy is 89.88%, the Kappa coefficient is 0.8026. Compared with the traditional BP neural network classification result, it is obviously find the CNN is more suitable for remote sensing image classification.