ISSN 1000-3665 CN 11-2202/P

    天山中段土地损毁自动分类模型的构建与应用

    Build and application of automatic classification model of land damage in the middle section of Tianshan Mountains

    • 摘要: 新疆天山中段矿产资源丰富,但高强度采矿活动导致土地损毁问题日益加剧。针对矿山土地损毁监测效率低、传统遥感解译依赖人工经验等问题,文章提出了基于神经网络的遥感影像土地损毁自动分类模型——SENetV2-COT-DeepLab V3+,该模型是在DeepLab V3+模型基础上,融合了Contextual Transformer(COT)模块与SENetV2模块,从而增强了上下文特征提取和通道注意力机制能力,优化模型对复杂矿山地物的分割能力。首先,根据高分系列遥感影像,构建了包含59198个样本的天山中段矿山样本集,通过数据增强扩展至177594个样本。然后,使用该数据训练SENetV2-COT-DeepLab V3+模型,提高其泛化能力与识别精度,精准掌握矿产资源开发造成的土地损毁分布和程度。最后,通过与FCN、UNeT、PSPNeT等模型进行对比试验得出该改进模型在平均交并比(MIoU)、平均召回率(mRecall)、平均精确率(mPrecision)和平均系数(mDice)等四项指标上均优于FCN、PSPNet等主流模型,分割精度较DeepLab V3+提升了1.63%~2.34%。基于该模型在pycharm平台搭建了矿区土地损毁类型深度学习遥感解译系统,目前该系统已部署至当地矿山管理部门,识别准确率达85%以上,实现了高精度、高效率的土地损毁识别,为矿区土地损毁动态监测与生态修复管理提供了智能化解决方案,推动矿山开发与环境保护的协调发展。

       

      Abstract: The middle section of the Tianshan Mountains in Xinjiang is rich in mineral resources, but the problem of land damage has been exacerbated by high-intensity mining activities. In order to solve the problems of low efficiency of mine land damage monitoring and traditional remote sensing interpretation relying on manual experience, this paper proposes an automatic classification model of remote sensing image land damage based on neural network-SENetV2-COT-DeepLab V3, which is based on the DeepLab V3 model and integrates the Contextual Transformer (COT) module and the SENetV2 module, so as to enhance the ability of context feature extraction and channel attention mechanism. Optimize the model's segmentation ability for complex mine features. Firstly, according to the high-resolution series of remote sensing images, a sample set of 59,198 samples was constructed for the middle section of the Tianshan Mountains, which was extended to 177594 samples through data augmentation. Then, the SENetV2-COT-DeepLab V3 model was trained to improve its generalization ability and recognition accuracy, and accurately grasp the distribution and extent of land damage caused by mineral resource development. Finally, through comparative experiments with FCN, UNeT, PSPNeT and other models, it is concluded that the improved model is better than the mainstream models such as FCN and PSPNet in four indicators: MIoU, mRecall, mPrecision, and mDice, and the segmentation accuracy is 1.63%~2.34% higher than that of DeepLab V3. Based on the model, a deep learning remote sensing interpretation system for land damage types in mining areas was built on the Pycharm platform, which has been deployed to the local mine management department, with a recognition accuracy of more than 85%, realizing high-precision and high-efficiency land damage identification, providing an intelligent solution for dynamic monitoring and ecological restoration management of land damage in mining areas, and promoting the coordinated development of mine development and environmental protection.

       

    /

    返回文章
    返回