林地在全球生态系统中扮演着至关重要的角色.但传统监督学习方法在林地提取上存在特征选择不精确与未能充分利用像元间的上下文关系等缺陷,导致林地提取精度不理想.针对上述问题,本文提出了一种基于改进PSPNet(Pyramid Scene Parsing Network)模型的高分辨率遥感影像林地提取方法.首先,利用高分二号遥感影像与全国第三次土地调查数据,制作高分辨率林地数据集.其次,通过在原始PSPNet模型的基础上加入SE(Squeeze and Excitation)注意力模块,改进PSPNet模型.实验结果表明,本文所改进的PSPNet模型的各项精度指标均优于其他方法,具有较高的提取精度.
Research on Forest Land Extraction Method Based on Improved PSPNet Model for High-resolution Remote Sensing Images
Woodlands play a crucial role in global ecosystems.However,traditional supervised learning methods have the defects of imprecise feature selection and failure to fully utilize the contextual relationship between image elements in forest land extraction,which leads to unsatisfactory accuracy of forest land extraction.To address the above problems,this paper proposes a forest land extraction method based on the improved PSPNet(Pyramid Scene Parsing Network)model for high-resolution remote sensing images.Firstly,the high-resolution forest land dataset is produced by utilizing the Gaofen-2 remote sensing image and the data of the third national land survey.Second,the PSPNet model is improved by adding the SE(Squeeze and Excitation)attention module to the original PSPNet model.The experimental results show that all the accuracy indexes of the improved PSPNet model in this paper are better than other methods and have high extraction accuracy.
deep learningPSPNetforest floor extractionSE attention mechanism
万方数据
维普
