CNN-based Large Scale Landsat Image Classification

Xuemei Zhao, Lianru Gao, Zhengchao Chen, Bing Zhang, Wenzhi Liao

Abstract: Large scale Landsat image classification is the key to acquire national even global land cover map. Traditional methods typically use only a small set of samples to train the classifier and result in unsatisfied classification results. To improve the performance of large scale Landsat image classification, we apply a convolutional neural network (CNN)-based method named PSPNet in this paper to learn spectral-spatial features from a large training set. By considering the complexities and the various sizes of objects captured in large scale Landsat images, PSPNet can utilize the global information as well as consider the targets with different sizes. In addition, the research area is oversampled with a small offset which can increase the amount of training samples in order to improve the performance of PSPNet on Landsat images. Moreover, PSPNet is finely tuned on the pretrained Resnet50. Experimental results show the efficiency of the CNN based methods for the large-scale land cover mapping. In particular, PSPNet can produce better results even than the provided reference land cover map, with overall accuracy reaching 83%.