NM-LinkNet: Cloud Detection from Remote Sensing Images with Non-local Operation and Multi-scale Feature Aggregation

Cloud detection is an important preprocessing process in remote sensing applications. Cloud detection methods have developed from traditional methods to deep learning methods which are widely used at present. However, current cloud detection methods still have limitations in the detection of thin clouds and broken clouds, mainly because the thin clouds are sparsely distributed and the size of broken clouds is relatively small. To solve the above difficult problems, this paper proposes a cloud detection network NM-LinkNet, which combines non-local operation and multi-scale feature aggregation, to improve the detection ability of thin clouds and broken clouds. Based on LinkNet50, NM-LinkNet uses non-local operation to obtain the long-distance context information of sparse distributed thin clouds to enhance the features of thin clouds. The multi-scale feature aggregation module designed in this paper is used to extract the features of clouds of different scales to improve the detection ability of small broken clouds. The experimental results on SPARCS dataset show that the IoU and F1 of NM-LinkNet proposed in this paper reach 87.50% and 93.33% respectively, which exceeds the other five mainstream deep learning methods in quantitative data and visualization results.