A Lightweight Network for Remote Sensing Image Cloud Detection

Based on the U-Net encoding and decoding structure, a more improved version of this model can be developed. An improved version of this encoder-decoder model has been developed based on the U-Net encoder-decoder structure. First of all, we replace all convolutions in the model with depthwise separable convolutions in the end-to-end training phase in order to reduce the total number of parameters and computations in the model. The second part of the work proposes the use of a Bottleneck Coordinate Attention Module (BCAM) in order to overcome the problems caused by depthwise separable convolutions being a source of accuracy reductions. By combining the advantages of both the Bottleneck Residual Block (BRB) and Coordinate Attention Mechanism (CAM), BCAM is able to extract deeper features in higher-dimensional space as well as embed coordinate information into the channel attention mechanism in order to extract more information over a wider area. As a result of the experimental results, it has been found that the parameters of the proposed method have declined by approximately 65% when compared to U-Net, and the proposed method achieves an intersection rate of 91.2% when compared to U-Net.