Learning Deep Features for Classification of Typical Ecological Environmental Elements in High-Resolution Remote Sensing Images

Abstract

Ecological environmental elements are greatly related to both humans and nature. The rapid development of remote sensing technology has provided us more and more high-resolution remote sensing images for monitoring these elements timely and objectively. However, it has been a great challenge to recognize these elements from such images due to their diversity and complexity. In this paper, a classification approach of ecological environmental elements based on deep learning features of objects is proposed. At first, a deep convolutional neural network (DCNN) is trained for discriminating different ecological environmental elements. To extract deep features of irregular-shaped regions, sub-images are clipped from each region and used to represent the corresponding region. Then, deep features of these sub-images are extracted by the trained DCNN. After that, the softmax classifier is used to predict class probabilities of all sub-images. The class of one region is determined by considering the class probabilities of its sub-images according to the winner-takes-all strategy. Finally, the thematic maps of ecological environmental elements are achieved. The proposed approach is evaluated by the classification experiments on a test set of typical ecological environmental elements in high-resolution remote sensing images, and the classification accuracy reaches to 98.44%. Moreover, the classification accuracy on irregular-shaped regions also reaches to 96.77%. These results have testified the effectiveness of the proposed approach.