Scholarlink

<p>Extracting wetland information is crucial for understanding wetland status, monitoring changes, and analysing resource distribution, thereby providing a scientific foundation for the sustainable development of wetland ecosystems. In order to improve the accuracy of wetland information extraction and reduce the problems of insufficient ability to extract multi-scale feature characteristics, difficulties in extracting small-scale patch categories and limited expression of heterogeneous boundary continuity in the process of wetland information extraction, this study introduces an improved network model based on DeepLabV3+. Firstly, ResNet-50 is adopted as the backbone network, and its deep residual structure is used to effectively solve the problem of insufficient feature extraction triggered by the shallow network MobileNetV2 in complex wetland scenarios, and improve the sensitivity of the model to wetland details and complex features. Secondly, the optimised Atrous Spatial Pyramid Pooling (ASPP) module expands the sensory field by introducing a cross-fertilisation mechanism in the parallel inflated convolutional branch, which is able to fuse the multi-scale information more efficiently, and significantly enhances the ability to identify the wetland boundaries and feature types, thus further improving the classification accuracy. Finally, the effect of the improved model was evaluated with the Liaohe River Estuary (LRE) wetland as the research object. The results showed that the network model’s overall accuracy (OA) was 97.6%, and the Kappa was 0.971. Compared to the original DeepLabV3 + as well as other network models (e.g., ERFNET, FCN, etc.), the OA of the models proposed in this study is improved by 0.3% to 8.7%, and the Kappa is improved by 0.004 to 0.106. Meanwhile, the accuracy of the network model proposed in this study is significantly enhanced in the extraction of feature types such as <i>Suaeda salsa</i>, <i>Phragmites australis</i>, and canal/ditch. The research results can be applied to finely extract wetland information in the LRE and provide reliable data support for wetland monitoring and management.</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

Wetland Information Extraction Method Based on Improved DeepLabV3 + in Liaohe River Estuary

  • 期刊论文

  • 发表日期：

    2026年1月14日

查看全文
  • Ying Wang,
  • Jinjie He,
  • Chang Wang,
  • Wen Zhang

摘要

Extracting wetland information is crucial for understanding wetland status, monitoring changes, and analysing resource distribution, thereby providing a scientific foundation for the sustainable development of wetland ecosystems. In order to improve the accuracy of wetland information extraction and reduce the problems of insufficient ability to extract multi-scale feature characteristics, difficulties in extracting small-scale patch categories and limited expression of heterogeneous boundary continuity in the process of wetland information extraction, this study introduces an improved network model based on DeepLabV3+. Firstly, ResNet-50 is adopted as the backbone network, and its deep residual structure is used to effectively solve the problem of insufficient feature extraction triggered by the shallow network MobileNetV2 in complex wetland scenarios, and improve the sensitivity of the model to wetland details and complex features. Secondly, the optimised Atrous Spatial Pyramid Pooling (ASPP) module expands the sensory field by introducing a cross-fertilisation mechanism in the parallel inflated convolutional branch, which is able to fuse the multi-scale information more efficiently, and significantly enhances the ability to identify the wetland boundaries and feature types, thus further improving the classification accuracy. Finally, the effect of the improved model was evaluated with the Liaohe River Estuary (LRE) wetland as the research object. The results showed that the network model’s overall accuracy (OA) was 97.6%, and the Kappa was 0.971. Compared to the original DeepLabV3 + as well as other network models (e.g., ERFNET, FCN, etc.), the OA of the models proposed in this study is improved by 0.3% to 8.7%, and the Kappa is improved by 0.004 to 0.106. Meanwhile, the accuracy of the network model proposed in this study is significantly enhanced in the extraction of feature types such as Suaeda salsa, Phragmites australis, and canal/ditch. The research results can be applied to finely extract wetland information in the LRE and provide reliable data support for wetland monitoring and management.