Seasonal-spatial distribution and prevailing wind analysis of Martian dust devils in Amazonis Planitia using a multi-scale attention network

The detection of dust devils on Mars poses significant challenges, primarily due to the substantial variability in target scales, the susceptibility of small-scale features to loss or distortion during feature extraction and fusion, and the interference from complex Martian backgrounds. To tackle these issues, we propose the Dynamic Triplet Fusion Attentive Net (DTFA-Net), a framework tailored for Martian dust devil detection. Within DTFA-Net, we design a Multi-Dimensional Dynamic Feature Pyramid Network (MDFPN), which is based on the Bi-directional Feature Pyramid Network (BiFPN) and enhances multi-level feature fusion by incorporating shallow-layer features and employing a cross-scale connection strategy. Additionally, we propose three innovative lightweight and plug-and-play modules: the Local-Channel Cross-Stage Module (LCCS) to boost feature diversity, the Progressive Feature Enhancement Module (PFE) to increase the focus on critical features, and the Triplet-Aware Cross-Stage Module (TACS) for capturing interactions across spatial and channel dimensions. Furthermore, the framework incorporates the Dynamic Head (DyHead), which uses multi-dimensional attention mechanisms to dynamically adjust to various scales, spatial positions, and detection challenges. Experimental results show that DTFA-Net achieved a detection Precision of 94.3%, a Recall of 92.8%, and mAP50 of 96.6% on the Amazonis Planitia dust devil dataset. Its overall performance significantly surpasses that of existing mainstream methods, while also demonstrating strong generalization capability on cross-regional datasets. Beyond detection, this framework was further applied to analyze the seasonal activity and spatial distribution patterns of Martian dust devils in Amazonis Planitia, and the prevailing winds of each season were examined to explore the mechanisms underlying the formation of activity hotspots. In addition, the model was extended to the ten core candidate landing sites of the Tianwen-3 mission to systematically assess dust devil distribution across these regions. Based on the detection results and previous studies, we suggest that four sites—Kasei Valles, Oxia Planum, McLaughlin Crater, and Mawrth Vallis—offer a more balanced trade-off among scientific value, dust-cleaning, and engineering safety, making them relatively ideal landing sites for the Tianwen-3 mission. Overall, this study provides important insights into the spatiotemporal distribution, activity patterns, and potential environmental risks of Martian dust devils, thereby offering valuable references for Mars exploration missions. Furthermore, it provides guidance for the optimized design and safe operation of spacecraft, and contributes scientific support for the planning and implementation of future Mars exploration endeavors.