Segmentation model based on convolutional neural networks for extracting vegetation from Gaofen-2 images

Abstract. Convolutional neural network (CNN) models achieve state-of-the-art performance for natural image semantic segmentation. An approach for extracting vegetation from Gaofen-2 (GF-2) remote sensing imagery based on the CNN model is presented. We constructed a convolutional encoder neural networks (CENN) consisting of two layers. The first layer has two sets of convolutional kernels for extracting the features of farmland and woodland, respectively. The second layer consists of two encoders that use nonlinear functions to encode the learned features and map the encoding results to the corresponding category number. In the training stage, samples of farmland, woodland, and other lands are categorically used to train the CENN. After training is accomplished, the CENN would acquire enough ability to accurately extract farmland and woodland from GF-2 imagery. The CENN was trained on 36 GF-2 images and tested on three other GF-2 images. We compared the proposed method to a deep belief network, a fully convolutional network, and a DeepLab model using the same images. The experiments demonstrate that the proposed approach improves upon the accuracy of existing approaches. The average precision, recall, and kappa coefficient of the proposed approach were 0.91, 0.87, and 0.86, respectively. Thus, the proposed approach is proven to effectively extract vegetation from GF-2 imagery.