Aerial Image Semantic Segmentation Using Spatial and Channel Attention

With the development of sensor, remote sensing has become an effective way to observe the Earth. Land cover classification is an important application of remote sensing data. Deep convolutional neural networks (DCNN) have good capability to extract effective image feature. In this paper, an automatic land cover classification method using spatial and channel attention was proposed. Encoder-Decoder structure was adopted as the network architecture. We used "channel attention" to enhance the function of effective channels. By using "spatial attention" to combining high-level semantic features with low-level features, better land objects boundaries were obtained. The proposed method was tested on ISPRS Potsdam 2D-Semantic Segmentation Challenge Dataset. The results showed that our proposed method, which uses spatial and channel attention with atrous spatial pyramid attention networks (SA-ASPA) outperforms the Deeplab_v3+ in terms of mean Intersection over Union (mIoU) for 5%.