Scholarlink

<p>Transfer learning has been applied to landslide susceptibility assessment to solve the problems of low predict accuracy and model instability caused by insufficient sample data. However, most research have made significant progress in exploring spatial transferability, the temporal dimension of landslide knowledge transfer remains largely uncharted. Besides, current ensemble-based transfer methodologies predominantly rely on conventional machine learning techniques. Therefore, this study proposes a deep learning-based stacking ensemble model to explore the transferability of landslide susceptibility in temporal dimension for Hong Kong. First, we preprocess study data in Hong Kong from 2008 to 2019, and divide them into source domain and target domain. Then, four deep learning architectures (Transformer, CNN, Attention-BiLSTM, and CNN-BiLSTM) are initially pre-trained on source domain data. Following this initialization phase, we fine-tuning these models with target domain data, and rank models for integration based on their transfer performance. Finally, we integrate the best two, three or four models and analyze the landslide susceptibility modelling results. Evaluations based on AUC measurements reveal that the transfer learning framework achieves enhanced predictive performance compared to deep learning implementations across most test scenarios. Furthermore, the deep learning-based ensemble framework enhances prediction accuracy by combining multiple transfer learning models, leveraging their complementary strengths to achieve superior performance. Among all transfer learning models, the integration of the four models yielded the best performance, with an AUC of 0.9091 in target domain from 2015 to 2017, which is 12.7% higher than Attention-BiLSTM model. The results indicate that the high susceptibility zones of Hong Kong are predominantly concentrated in the eastern regions and Lantau Island, while showing an emerging spatial shift toward central territories. This study explore the transferability of landslide knowledge across time dimension, provide a novel perspective for spatiotemporal landslide susceptibility assessment in data-scare temporal.</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

Spatiotemporal landslide susceptibility modeling based on integrated transfer learning

  • 期刊论文

  • 发表日期：

    2025年9月27日

查看全文
  • Zhengshan Tian,
  • Yi Wang,
  • Zhice Fang,
  • Yanyan Wang,
  • Baoyu Du

摘要

Transfer learning has been applied to landslide susceptibility assessment to solve the problems of low predict accuracy and model instability caused by insufficient sample data. However, most research have made significant progress in exploring spatial transferability, the temporal dimension of landslide knowledge transfer remains largely uncharted. Besides, current ensemble-based transfer methodologies predominantly rely on conventional machine learning techniques. Therefore, this study proposes a deep learning-based stacking ensemble model to explore the transferability of landslide susceptibility in temporal dimension for Hong Kong. First, we preprocess study data in Hong Kong from 2008 to 2019, and divide them into source domain and target domain. Then, four deep learning architectures (Transformer, CNN, Attention-BiLSTM, and CNN-BiLSTM) are initially pre-trained on source domain data. Following this initialization phase, we fine-tuning these models with target domain data, and rank models for integration based on their transfer performance. Finally, we integrate the best two, three or four models and analyze the landslide susceptibility modelling results. Evaluations based on AUC measurements reveal that the transfer learning framework achieves enhanced predictive performance compared to deep learning implementations across most test scenarios. Furthermore, the deep learning-based ensemble framework enhances prediction accuracy by combining multiple transfer learning models, leveraging their complementary strengths to achieve superior performance. Among all transfer learning models, the integration of the four models yielded the best performance, with an AUC of 0.9091 in target domain from 2015 to 2017, which is 12.7% higher than Attention-BiLSTM model. The results indicate that the high susceptibility zones of Hong Kong are predominantly concentrated in the eastern regions and Lantau Island, while showing an emerging spatial shift toward central territories. This study explore the transferability of landslide knowledge across time dimension, provide a novel perspective for spatiotemporal landslide susceptibility assessment in data-scare temporal.