Xu Wu

Automated analysis of UAV imagery is crucial for remote sensing but faces challenges like extreme scale variations, complex background clutter, and the vanishing of fine-grained details. To address these limitations, we propose MMAF-YOLO, a novel object detection network based on multi-scale mutual aggregation and fine-grained feature modulation. First, we design a Multi-Scale Mutual Aggregation Fusion (MMAF) block in the backbone to replace the standard C3k2 module. This effectively captures complementary multi-scale information and mitigates feature loss for small targets. Second, we devise a Fine-Grained Feature Modulation Neck incorporating a Modulation Fusion Module (MFM) and an Efficient Up-Convolution Block with Channel Shuffle (EUCB-CS) to resolve semantic misalignment and repair structural distortions during upsampling, supplemented by a high-resolution P2 detection head. Third, a joint optimization strategy combining Inner-IoU and Scale-based Dynamic Loss (SDLoss) is employed to stabilize regression and improve recall. To satisfy diverse deployment requirements, we develop a scalable model series consisting of four variants (n, s, m, l). Comprehensive evaluations conducted on the VisDrone2019 dataset demonstrate that MMAF-YOLO surpasses the YOLO11 baseline across all model scales. Specifically, the MMAF-YOLO-s variant improves mAP@0.5 by 5.0% over YOLO11-s with reduced parameters. Furthermore, evaluations on the HIT-UAV (infrared) and HazyDet (adverse weather) datasets confirm the model’s robust generalization capabilities across diverse remote sensing modalities and extreme environmental conditions.

In recent years, remote sensing technology has been extensively used in detecting and managing natural disasters, playing a vital role in the early identification of events like landslides. The integration of deep learning models has considerably enhanced the efficiency and accuracy of landslide detection particularly in automating the analysis and quickly identifying affected areas. However, existing models often face challenges, such as incomplete feature extraction, loss of contextual information, and high computational complexity. To overcome these challenges, we propose an innovative landslide detection model, Lights-Transformer, which is designed to improve both the accuracy and efficiency. This model employs an encoder–decoder architecture that incorporates multi-scale contextual information and an efficient attention mechanism, effectively capturing both local and global features of images while minimizing information loss. By introducing a Fusion Block for enhanced multi-angle feature fusion and a Light Segmentation Head to boost inference speed, Lights-Transformer extracts detailed feature maps from high-resolution remote sensing images, enabling the accurate identification of landslide regions and significantly improving detection accuracy. Compared to existing state-of-the-art landslide detection models, Lights-Transformer offers considerable advantages in accuracy, precision, and computational efficiency. On the GDCLD dataset, Lights-Transformer achieves an mIoU of 85.11%, accuracy of 97.44%, F1 score of 91.49%, kappa value of 82.98%, precision of 91.46%, and recall of 91.52%, demonstrating its exceptional performance.

Landslides, as a prevalent geological hazard, pose a severe threat to both the environment and human society. The rapid and accurate identification of landslide-prone areas is crucial for disaster response, risk assessment, and urban planning. This paper proposes a novel deep learning-based landslide detection method ResM-FusionNet, which leverages ResNet-50 as the backbone for feature extraction and integrates a multi-layer perceptron as the decoder to enhance segmentation accuracy. To address the challenges of complex terrain and boundary detail detection in landslide-prone regions, we introduce a novel loss function, RLoss, which integrates masking mechanisms, semantic weighting, and a Top-K pixel loss averaging strategy. Experimental results on remote sensing datasets demonstrate that ResM-FusionNet significantly outperforms existing models. Specifically, ResM-FusionNet achieves 94.33% accuracy, 85.73% F1-score, and a Kappa coefficient of 70.12%, surpassing other models (e.g., SegFormer, DeepLabv3, and UNet) by 4.4%, 7.7%, and 17.6% in accuracy, respectively. Moreover, ResM-FusionNet excels in boundary detection, achieving an IoU of 0.7545, precision of 85.61%, and recall of 83.92%. These findings indicate that the proposed method not only provides robust and accurate landslide detection but also enhances segmentation performance in complex terrains.