Scholarlink

<p>Landslides are natural calamities impacting the environments worldwide and prove to be a cataclysmic disaster for human settlements, animal habitats and infrastructures. Mitigating its life-threatening risks through preventive evacuations demands timely detection in remotely sensed complex satellite images, even when the landslide regions appear tiny. Hence, this paper presents a novel encoder-decoder architecture, MRF-LaD, employing multi-residual convolutional layers along the encoder, decoder and the bottleneck paths with skip connections for automated landslide detection in satellite images. The architecture also integrates light-weight Efficient channel attention network between the encoder layers for focused feature weighting, along with a strategically placed parallel cross-stage partial network to refine multi-scale and multi-resolution features. The bottleneck incorporates a learned single-stage transformer block for capturing and propagating fine image details during the segmentation mask construction. The proposed MRF-LaD architecture attained effective segmentation for the benchmark Landslide4sense dataset (Ghorbanzadeh in 15:9927–9942,&#xa0;2022;&#xa0;Ghorbanzadeh in&#xa0;Landslide4sense: Reference benchmark data and deep learning models for landslide detection,&#xa0;2022) with an <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(F1=70\%\)</EquationSource> </InlineEquation> and <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(mIoU=78\%\)</EquationSource> </InlineEquation> approximately, outperforming 10 baseline and 15 state-of-the-art existing techniques. The efficient landslide segmentation in images covering varied terrains and multiple regions worldwide claims model’s strong generalization capability, thereby contributing to multiple United Nation’s Sustainable Development Goals.</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

Mrf-lad: a hybrid transformer-augmented multi-residual framework with efficient channel attention and cross-stage partial network for landslide detection

  • 期刊论文

  • 发表日期：

    2025年12月8日

查看全文
  • Srinidhi Kannan,
  • Rimjhim Padam Singh

摘要

Landslides are natural calamities impacting the environments worldwide and prove to be a cataclysmic disaster for human settlements, animal habitats and infrastructures. Mitigating its life-threatening risks through preventive evacuations demands timely detection in remotely sensed complex satellite images, even when the landslide regions appear tiny. Hence, this paper presents a novel encoder-decoder architecture, MRF-LaD, employing multi-residual convolutional layers along the encoder, decoder and the bottleneck paths with skip connections for automated landslide detection in satellite images. The architecture also integrates light-weight Efficient channel attention network between the encoder layers for focused feature weighting, along with a strategically placed parallel cross-stage partial network to refine multi-scale and multi-resolution features. The bottleneck incorporates a learned single-stage transformer block for capturing and propagating fine image details during the segmentation mask construction. The proposed MRF-LaD architecture attained effective segmentation for the benchmark Landslide4sense dataset (Ghorbanzadeh in 15:9927–9942, 2022; Ghorbanzadeh in Landslide4sense: Reference benchmark data and deep learning models for landslide detection, 2022) with an \(F1=70\%\) and \(mIoU=78\%\) approximately, outperforming 10 baseline and 15 state-of-the-art existing techniques. The efficient landslide segmentation in images covering varied terrains and multiple regions worldwide claims model’s strong generalization capability, thereby contributing to multiple United Nation’s Sustainable Development Goals.