Optical Remote Sensing Image Cloud Detection with Self-Attention and Spatial Pyramid Pooling Fusion

Cloud detection is a key step in optical remote sensing image processing, and the cloud-free of great significance for land use classification, change detection, long time-series landcover monitoring. Traditional cloud methods based on spectral texture features have acquired certain effects complex scenarios, such as cloud–snow mixing, but there still large room improvement terms generation ability. In recent years, with deep-learning has significantly improved accuracy regions high-brightness feature mixing areas. However, existing deep learning-based limitations. For instance, few omission alarms commission exist edge regions. At present, learning are gradually converted from pure convolutional structure to global extraction perspective, attention modules, computational burden also increased, which difficult meet rapidly developing time-sensitive tasks, onboard real-time imagery. To address above problems, this manuscript proposes high-precision network fusing self-attention module spatial pyramidal pooling. Firstly, we DenseNet backbone, then semantic extracted by combining pyramid pooling module. Secondly, solve problem unbalanced training samples, design weighted cross-entropy loss function optimize it. Finally, assessed. With quantitative comparison experiments different images, Landsat8, Landsat9, GF-2, Beijing-2, results indicate that, compared feature-based methods, can effectively distinguish confusion-prone region using only visible three-channel reduces number required bands. Compared other at higher overall efficiency relatively optimal.