Scholarlink

<p>With the widespread application of remote sensing imagery, deep learning-based landslide detection methods have emerged as a new direction in technological development. However, existing deep learning approaches still face challenges of insufficient accuracy or suboptimal computational efficiency when addressing complex terrain conditions and diverse landslide morphologies. This study proposes a Transformer-based landslide detection model, named Complex Terrain Landslide DEtection TRansformer (CTL-DETR), which integrates Synergistic Multi-Attention Feature Extraction (C2f-SMAG), Mixed Aggregation and Faster Fusion (MA-Fusion), and Context-Guided Downsampling (CG-Down) modules. The C2f-SMAG module combines channel, spatial, pixel, and multi-head self-attention mechanisms to enhance the model’s ability to capture global dependencies and local details. The MA-Fusion module employs a multi-branch architecture, integrating bypass convolution, depthwise separable convolution, and partial convolution, to achieve efficient integration of channel and spatial features. The CG-Down module combines local features with global contextual information to optimize the integrity of feature transmission. This study constructs a Diverse Landslide Dataset (DLD), on which CTL-DETR achieved an F1 score of 75.8% and an mAP50 of 78.7%. Compared with mainstream object detection models and existing landslide detection methods, CTL-DETR demonstrates superior performance in terms of precision, recall, and other metrics, validating the effectiveness of the proposed approach. Further generalization experiments demonstrate the model’s robustness in cross-regional landslide detection, with mAP50 scores of 74.7% and 91.2% on the Bijie and Luding datasets, respectively. Additionally, CTL-DETR maintains relatively low computational complexity and parameter count, making it more suitable for wide-scale landslide detection applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

CTL-DETR: a landslide detection algorithm for complex terrains

  • 期刊论文

  • 发表日期：

    2025年10月11日

查看全文
  • Xiaopeng Zang,
  • Jiajun Li,
  • Guoqing Ma,
  • Momo Zhi,
  • Shitong Chen

摘要

With the widespread application of remote sensing imagery, deep learning-based landslide detection methods have emerged as a new direction in technological development. However, existing deep learning approaches still face challenges of insufficient accuracy or suboptimal computational efficiency when addressing complex terrain conditions and diverse landslide morphologies. This study proposes a Transformer-based landslide detection model, named Complex Terrain Landslide DEtection TRansformer (CTL-DETR), which integrates Synergistic Multi-Attention Feature Extraction (C2f-SMAG), Mixed Aggregation and Faster Fusion (MA-Fusion), and Context-Guided Downsampling (CG-Down) modules. The C2f-SMAG module combines channel, spatial, pixel, and multi-head self-attention mechanisms to enhance the model’s ability to capture global dependencies and local details. The MA-Fusion module employs a multi-branch architecture, integrating bypass convolution, depthwise separable convolution, and partial convolution, to achieve efficient integration of channel and spatial features. The CG-Down module combines local features with global contextual information to optimize the integrity of feature transmission. This study constructs a Diverse Landslide Dataset (DLD), on which CTL-DETR achieved an F1 score of 75.8% and an mAP50 of 78.7%. Compared with mainstream object detection models and existing landslide detection methods, CTL-DETR demonstrates superior performance in terms of precision, recall, and other metrics, validating the effectiveness of the proposed approach. Further generalization experiments demonstrate the model’s robustness in cross-regional landslide detection, with mAP50 scores of 74.7% and 91.2% on the Bijie and Luding datasets, respectively. Additionally, CTL-DETR maintains relatively low computational complexity and parameter count, making it more suitable for wide-scale landslide detection applications.