摘要
若尔盖湿地作为中国最大的高原沼泽湿地分布区,具有丰富的野生动植物资源和重要的生态功能。然而,湿地景观组分复杂且分布零散,传统的分类方法在处理这类数据时面临挑战。为了解决此问题,笔者提出了一种Trans-MLP模型,实现湿地景观组分提取。首先,从遥感数据提取空间特征,以丰富数据特征信息;其次,设计了一个由Transformer模块和多层感知机结构组成的网络模型;最后,将提取的光谱和空间特征信息输入模型,完成对若尔盖湿地像素级的分类。研究表明:2021年与2023年若尔盖湿地实验数据总体精度分别为99.1%和99.2%,相比于仅使用单一的光谱特征提升了约2%,且分类结果更符合地物真实情况;选择的两种对比模型VGG19和ResNet50模型的总体精度在95.9%~96.5%之间,笔者所提分类模型总体效果优于对比模型。综上所述,结合光谱-空间特征输入的Trans-MLP模型在若尔盖湿地景观组分提取中具有较高的应用潜力,为湿地保护和管理提供了有力的技术支持。
Zoige Wetland is the largest plateau swamp wetland distribution area in China,which has rich wildlife resources and important ecological functions.However,wetland landscape components are complex and scattered,making traditional classification methods challenging to apply to such data.To address this problem,this paper proposes a Trans-MLP model to extract components of the wetland landscape.First,spatial features are extracted from remote sensing data to enrich feature information;second,a network model consisting of a Transformer module and an MLP structure is designed;finally,the extracted spectral and spatial features are input into the model to complete pixel-level classification of the Zoige wetland.Research shows that the overall accuracy of the Zoige wetland experimental data for 2021 and 2023 is 99.1%and 99.2%,respectively,representing an improvement of approximately 2%compared to using only spectral features.Additionally,the classification results better align with the actual landscape.In contrast,the selected comparison models,VGG19 and ResNet50,achieved overall accuracies of 95.9%-96.5%,while the proposed classification model outperformed them.In summary,the Trans-MLP model,with combined spectral-spatial feature input,shows high potential for application in extracting landscape components in the Zoige wetland,providing robust technical support for wetland conservation and management.
作者
蒋烨
陈建华
王炳乾
左世祥
崔富强
张晓锋
张倩
JIANG Ye;CHEN Jianhua;WANG Bingqian;ZUO Shixiang;CUI Fuqiang;ZHANG Xiaofeng;ZHANG Qian(College of Geophysics,Chengdu University of Technology,Chengdu 610059,China;Aba Prefecture Institute of Natural Resources and Scientific and Technological Information,Wenchuan 623000,China)
出处
《物探化探计算技术》
2026年第1期1-14,共14页
Computing Techniques For Geophysical and Geochemical Exploration
基金
四川省重点研发项目(2024YFTX0004)
阿坝州科技计划项目(R23YYJSYJ0006)。
关键词
若尔盖湿地
景观组分
空间特征
多头自注意力
多层感知机
Zoige Wetland
landscape components
spatial features
multi-head self-attention
multilayer perceptron
