The coherent speckle noise in synthetic aperture radar (SAR) imagery, together with complex sea clutter and large variations in ship target scales, poses significant challenges to accurate and robust ship detection, particularly under strict lightweight constraints required by satellite-borne and airborne platforms. To address this issue, this paper proposes a high-precision lightweight detection network, termed High-Lightweight Net (HLNet), specifically designed for SAR ship detection. The network incorporates a novel multi-scale backbone, Multi-Scale Net (MSNet), which integrates dynamic feature completion and multi-core parallel convolutions to alleviate small-target feature loss and suppress background interference. To further enhance multi-scale feature fusion while reducing model complexity, a lightweight path aggregation feature pyramid network, High-Lightweight Feature Pyramid (HLPAFPN), is introduced by reconstructing fusion pathways and removing redundant channels. In addition, a lightweight detection head, High-Lightweight Head (HLHead), is designed by combining grouped convolutions with distribution focal loss to improve localization robustness under low signal-to-noise ratio conditions. Extensive experiments conducted on the public SSDD and HRSID datasets demonstrate that HLNet achieves mAP50 scores of 98.3% and 91.7%, respectively, with only 0.66 M parameters. Extensive evaluations on the more challenging CSID subset, composed of complex scenes selected from SSDD and HRSID, demonstrate that HLNet attains an mAP50 of 75.9%, outperforming the baseline by 4.3%. These results indicate that HLNet achieves an effective balance between detection accuracy and computational efficiency, making it well-suited for deployment on resource-constrained SAR platforms.
Fire occurrence in Southwest China is jointly shaped by meteorological conditions, topography, vegetation status, and human activities. To improve the interpretability and validation rigor of regional fire susceptibility assessment, this study developed a grid-day-based susceptibility assessment framework for Yunnan, Sichuan, Guizhou, and Chongqing using MODIS active-fire detections and multi-source environmental data from 2015 to 2024 at a 5 km grid resolution. A sensitivity analysis was conducted to determine the training sample configuration, and a 1:2 positive-to-negative sampling ratio was adopted. Random Forest (RF), Support Vector Machine (SVM), and Logistic Regression (LR) were compared, and SHapley Additive exPlanations (SHAP), together with partial dependence plots (PDP), were used to interpret key drivers and their interactions. Data from 2015 to 2018 were used for model training, while data from 2019 to 2024 were used to evaluate the model’s cross-year transferability within the same study domain, rather than full spatiotemporal independence. The results show that the 1:2 sampling ratio achieved a favorable balance between fire detection and false-alarm control. In five-fold stratified cross-validation, RF outperformed LR and SVM (AUC = 0.9167; F1-score = 76.70%). In the cross-year transferability test, areas classified as high and very high susceptibility captured 62.04–68.95% of the observed fire points while accounting for less than 32% of the total area. Soil moisture and maximum temperature contributed most strongly to the model output, and their interaction revealed a pronounced dry-hot statistical response pattern associated with elevated susceptibility. Fire susceptibility also exhibited stable positive spatial autocorrelation, with hotspot areas concentrated in the dry-hot valleys near the Sichuan-Yunnan border and in central-southern Yunnan. Because the model was built with under-sampled negatives and same-day environmental matching, the output should be interpreted as a relative fire susceptibility index for spatial assessment and statistical attribution rather than as a calibrated occurrence probability or a forward-looking daily forecast.