CEDA-UNet: computational-efficient dilated attention UNet for multiclass cloud image segmentation

Image segmentation technology based on Convolutional Neural Network (CNN) to process cloud images taken by satellite remote sensing has become a valuable research field. However, in previous practice, the shallow network of the traditional UNet model has been found to be inadequate for dealing with satellite remote sensing data that lacks contextual information. Additionally, it suffers from information bottlenecks if the hopping connections are not optimised. This paper designs and improves the UNet architecture as Computational-Efficient Dilated Attention UNet (CEDA-UNet). Specifically, we initially proposed a method to introduce the attention mechanism into the decoder, and subsequently employed the dilated convolution method to replace the conventional convolution. The model can be competent for image segmentation tasks of extracting contextual information. In a dataset named 'Understanding Clouds from Satellite Images' comprising satellite remote sensing images, cloud patterns are manually annotated by expert meteorologists. This dataset allows for the evaluation of the cloud segmentation detection effect of the CEDA-UNet model in comparison to five other classic models, including Attention UNet, U2-Net, UNet, UNet3+ and VNet.