Improving Cloud Detection over Bright Underlying Surfaces Using a Novel MBUNet Framework for Remote Sensing Imagery

Accurate cloud detection is a prerequisite for remote sensing image applications. Cloud detection over bright surfaces is always a complex problem. Because of the similar spectral characteristics of clouds and bright surfaces, it is challenging to identify clouds from bright surfaces accurately. A novel MBUNet framework is proposed to improve cloud detection in remote sensing images of bright underlying surfaces. The network uses Bi-directional convolutional LSTM (BConvLSTM) to obtain multi-scale feature information by a nonlinear combination of low and high-dimensional image features, extract global features, and reduce information loss. Multi-scale context attention module (MCAM) enables the network to focus on the cloud region, suppress the non-cloud region, and learn delicate high-dimensional semantic information in conjunction with context information, thus improving the expressing ability characteristics and getting accurate cloud mask. In addition, a new loss function is used to focus on foreground information and boundary information. Cloud detection experiments were carried out on Landsat 8 Biome and GF-1WFV datasets, respectively, and the results show that the method performs well in cloud detection over bright surfaces, effectively reducing the misjudgment between clouds and bright surface objects.