Bisupervised network with pyramid pooling module for land cover classification of satellite remote sensing imagery

Land cover classification is a significant task in remote sensing that aims at land cover monitoring and adjustment. Instance segmentation in deep learning has been widely used in land cover classification. However, this method requires high quality, labor-intensive pixel-level annotations. A bisupervised pipeline is proposed especially for datasets with ambiguous samples, which reduces the requirement of labeling accuracy. In order to supervise learning with ambiguous annotation samples, the pipeline has two losses to feedback, namely main loss and auxiliary loss. The main loss is still responsible for the expected output. According to the difference degree between ambiguous samples, categories are constructed artificially. The auxiliary loss generates feedback through categories, sharing decoder layers of main output during the training process. In the prediction process, auxiliary loss is used to update the main loss results. The neural network adopts the structure of U-Net with the pyramid pooling module, in which the multiscale feature is used in the feature extraction process. We also compare five different backbones and choose Inception-V3 as the backbone to improve the feature extraction capabilities of network encoders. The use of transposed convolution instead of traditional upsampling in decoders can improve the segmentation details. Our bisupervise network obtains Jaccard index of 83.936% and F 1-score of 91.17% on Gaofen-2 imagery dataset. The results demonstrate that the proposed method can achieve better classification performance in datasets with ambiguous annotations.