GACraterNet: A collaborative geometry-attribute domain network for enhanced detection of Martian impact craters

Abstract

Accurately understanding the local and global distribution, categories and morphological parameters of impact craters on Mars, including variations across the southern highlands, northern lowlands, equatorial region and polar zones, is crucial for revealing the geological history and environmental changes. To this end, this paper proposes a multi-task deep learning framework, GACraterNet(Geometric and Attribute Domain-based Crater Detection Network), which addresses both impact crater detection and attribute extraction while facilitating mutual enhancement between these tasks. GACraterNet comprises two main components: the geometric-domain module and the attribute-domain module. The geometric-domain module features a detection network named M${\text{S}}^2$CraterNet, which encodes crater features from two data sources: digital elevation models (DEMs) and digital orthophoto maps (DOMs) using a dual backbone network. This module incorporates a multi-source, multi scale feature fusion module (M${\text{S}}^2$-FFM) to integrate the features, enabling the detection of craters larger than 1 km in diameter. The attribute-domain module is designed to perform three tasks: segmentation, classification and extraction of morphological parameters. First, Segment Anything Model (SAM) is utilized for unsupervised semantic segmentation on terrain maps within the bounding boxes predicted by M${\text{S}}^2$CraterNet. This step enhances the extraction of crater foregrounds and optimizes the positioning and sizing of the bounding boxes. The resultant crater foregrounds are then input to the Swin Transformer network, which categorizes craters into four types: bowl-shaped, flat floor, central peak and central pit. Finally, radial profiles of each crater type are analyzed to extract their 2.5D morphological parameters, followed by a comparative analysis of the morphological differences among the various categories. Validation results on the HRSC Mars remote sensing dataset indicate that M${\text{S}}^2$CraterNet achieved a mean Average Precision (mAP50) of 79.4%, with precision and recall rates of 78.4% and 73.3%, respectively. These results significantly outperform the detection results obtained from a single data source. Furthermore, Swin Transformer attained an overall classification accuracy of 83.9% for the craters, with specific classification F1-score for bowl-shaped, central peak, central pit and flat floor craters reaching 91.5%, 83.4%, 35.3% and 71.9%, respectively. The source code of our GACraterNet is available at https://github.com/shincccc/GACraterNet.