Abstract: The classification of high-resolution remote sensing image (HSRI) based on deep convolutional neural network (DCNN) is one of the hotspots in remote sensing intelligent interpretation technology. However, existing classification networks rarely exploited the collaborative effect of multi-features, and cannot capture the complex spatial relationships of ground objects in HSRI. In order to further utilize spatial information and improve classification accuracy, we propose a multi-features collaborative deep network (MFCDN). In the proposed method, multi-type shallow features are first extracted as the input of MFCDN. Then, the multi-scale feature extraction module is utilized to extract the feature information of different spatial scales. Next, after dynamic weighting by channel and spatial attention mechanisms, multiple feature extraction layers and down-sampling layers are used to extract semantic information and perform feature fusion by element-by-element addition. Finally, the classification results are obtained through the multilayer perceptron combined with the Softmax function. Experimental results demonstrate that the proposed MFCDN outperforms other related methods in teams of classification accuracy and generalization performance.

Key words: high-resolution remote sensing image, classification, multi-feature collaboration, deep convolutional neural networks, attention mechanism