FERA-Net: A Novel Algorithm for Mars Water-Ice Cloud Segmentation Integrating Feature Enhancement, Residual, and Attention Mechanisms

This study proposes the Feature-Enhanced Residual Attention U-Net (FERA-Net) model, which is trained using cloud mask label data generated by the semiautomatic CM7 method. The aim is to enhance the automation of water-ice cloud segmentation tasks and improve the segmentation accuracy based on existing deep learning methods. Water ice clouds are a crucial component of Martian weather phenomena, and their precise description is important for understanding Martian climate changes. However, the Martian surface topography's complexity and the atmospheric environment's dynamic changes make accurate identification of water ice clouds difficult. To tackle this, we combine feature channels, spatial-channel attention bridge (SC), attention gating (AG), and a double convolution residual block (R2CBL) to enhance feature extraction and gradient stability. We introduce a new feature channel by calculating the ratio of the blue (B) and red (R) channels(B/R) to enhance the expression of cloud features. The SC, incorporating spatial attention blocks (SAB) and channel attention blocks (CAB), along with the AG, improves the model's ability to capture key features. The R2CBL, through double convolution operations and residual connections, addresses the gradient vanishing problem and enhances feature extraction. Experimental results demonstrate that, compared to existing deep learning methods, FERA-Net achieves significant improvements in the task of Martian water-ice cloud segmentation. Specifically, the F1-score increases from 75.16% to 89.73% (an improvement of 19.4%), and the Jaccard index improves from 78.93% to 90.01% (an increase of 14.1%).This research provides theoretical support for in-depth exploration of the Martian atmospheric environment and water cycle, and it can assist in planning and executing future Mars exploration missions.