基于机器学习的高分辨率遥感影像地理信息提取方法  

A machine learning-based method for geographic information extraction from high-resolution remote sensing imagery

作　　者：王文龙 王可可 WANG Wenlong;WANG Keke(Zhejiang Huahui Geotechnical Survey Company Limited,Shaoxing,Zhejiang 312000,China;Zhejiang Provincial Natural Resources Group Spatial Information Company Limited,Shaoxing,Zhejiang 312000,China)

机构地区：[1]浙江华汇岩土勘测有限公司,浙江绍兴312000 [2]浙江省自然资源集团空间信息有限公司,浙江绍兴312000

出　　处：《北京测绘》2025年第4期528-533,共6页Beijing Surveying and Mapping

基　　金：国家自然科学基金(42261074)。

摘　　要：为保证高分辨率遥感影像信息提取结果的完整性,本文提出一种基于机器学习的高分辨率遥感影像地理信息提取方法。该方法采用色调、亮度和饱和度(HIS)空间变换技术,变换高分辨率地理遥感影像颜色空间,提升影像中目标的可分性;将变换后的地理影像输入卷积注意力机制网络模型中,结合空间和通道两种注意力机制,生成特征提取模块,提取并融合影像特征,从而获取特征图;解码层采用级联条件随机场(CRF)逐层处理特征图,增强上下层特征之间的关联,保证特征的丰富性,进行特征图边缘轮廓边界的优化,提升特征图的空间信息,获取地理信息提取结果。测试结果显示:变换处理后,影像的对比度和平均梯度分别在0.933和0.942以上,说明该方法较好地保证影像的变换质量,能够完整提取影像中的目标信息。To ensure the completeness of geographic information extraction from high-resolution remote sensing imagery,this paper proposed a machine learning-based method for geographic information extraction from high-resolution remote sensing images.The method utilizes the hue,intensity,and saturation(HIS) color space transformation technique to transform the color space of high-resolution remote sensing imagery,enhancing the separability of targets within the image.The transformed geographic image was then input into a convolutional attention mechanism network model,which integrated spatial and channel attention mechanisms to create a feature extraction module.This module extracted and fused image features,resulting in a feature map.The decoding layer uses a cascaded conditional random field(CRF) to process the feature map layer by layer,enhancing the correlation between features at different layers,ensuring feature richness,optimizing edge contours in the feature map,and improving spatial information to obtain the final geographic information extraction results.Test results show that after transformation,the image's contrast and average gradient are both above 0.933 and 0.942,respectively,indicating that the method ensures good transformation quality and can effectively extract target information from the image.

关 键 词：机器学习 地理信息提取 特征图 级联条件随机场(CRF) 色调、亮度和饱和度(HIS)空间变换 

分 类 号：P227[天文地球—大地测量学与测量工程]