Fisher可分比和递归特征消除算法优化机器学习的不透水面材质提取  

Optimized Machine Learning based on Fisher Discriminant Ratio and Recursive Feature Elimination Algorithms for Material Extraction of Impervious Surface

作　　者：徐红涛 何斌 杨宏 朱文泉 贺相綦 郝坤钰 XU Hongtao;HE Bin;YANG Hong;ZHU Wenquan;HEXiangqi;HAO Kunyu(State Key Laboratory of Earth Surface Processes and Resource Ecology,Faculty of Geographical Science,Beijing Normal University,Beijing 100875,China;Akesu National Station of Observation and Research for Oasis Agro-ecosystem,Aksu 843000,China;Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing 100094,China;School of Electronics and Information,Northwestern Polytechnical University,Xi’an 710129,China)

机构地区：[1]北京师范大学地理科学学部地表过程与资源生态国家重点实验室,北京100875 [2]新疆阿克苏绿洲农业生态系统国家野外观测研究站,新疆阿克苏843000 [3]中国科学院空天信息创新研究院,北京100090 [4]西北工业大学电子信息学院,陕西西安710129

出　　处：《遥感技术与应用》2024年第6期1466-1477,共12页Remote Sensing Technology and Application

基　　金：高分专项航空观测系统应用校飞与验证项目(30-H30C01-9004-19/21);第三次新疆综合科学考察项目(2022xjkk0106);环境演变与自然灾害教育部重点实验室开放课题(2024-KF-11)。

摘　　要：不透水面作为城市地表要素的重要组成成分,了解其空间分布可为城市的发展以及灾害防护提供科学的参考依据。然而由于光谱的相似性,准确获取不透水面材质具有一定的挑战性。本研究借助于亚米级航空可见光遥感数据,通过构建光谱、指数、纹理和形状等对象特征,使用Fisher可分比(Fisher Discriminant Ratio,FDR)和递归特征消除算法(Recursive Feature Elimination,RFE)优选最终的建模变量,结合随机森林(RandomForest,RF)、XGBoost、GBDT、CatBoost、LightGBM等算法分别构建了不同不透水层材质的分类模型(FDR-RFE-RF,FDR-RFE-XGBoost,FDR-RFE-GBDT,FDR-RFE-CatBoost,FDR-RFE-LightGBM),选取最优模型绘制了区域不透水面材质的空间分布。结果表明:除GBDT和LightGBM外,使用经FDR和RFE算法优选的变量构建的模型在点尺度上总体精度和Kappa系数分别提高了0.933%~1.171%和1.229%~1.542%,并改善了分类结果空间破碎的现象。结合点尺度的验证精度和空间上的分类结果发现FDR-RFE-RF的性能最优(OA=0.926,Kappa Coefficient=0.906),且其提取的研究区的不透水面材质的分布特征基本与实际情况一致。研究结果表明变量优选在一定程度上可以提升基于机器学习的不透水面材质提取的鲁棒性,同时也验证了基于亚米级航空可见光影像在城市不透水层材质提取的可行性。Impervious surface is an important component of urban surface elements.Knowledge about its spatial distribution can provide a scientific reference for urban development and disaster protection.However,due to the similarity of spectra,it is challenging to accurately obtain the impermeable surface material.Object-based and machine learning methods are applied to extract materials of urban impervious.Based on the aerial visible waveband remote sensing imagery with a spatial resolution of sub-meter,the variables including spectrum,vegetation index,texture and shape properties are constructed.Combining Fisher Discriminant Ratio(FDR)and Recursive Feature Elimination(RFE)algorithms,the final variables for training machine learning model were determined.Machine learning algorithms such as Random Forest(RF),XGBoost,GBDT,CatBoost and LightGBM were developed to construct impervious material classification models(FDR-RFE-RF,FDR-RFE-XGBoost,FDR-RFE-GBDT,FDR-RFE-CatBoost,FDR-RFE-LightGBM).The best model was selected and to extract the spatial distribution of impervious materials in the study area by comparing the accuracy and the local spatial pattern of impervious materials of different models.The results showed that,compared with the impervious surface material extraction model constructed using all variables,except for GBDT and LightGBM,the overall accuracy and Kappa coefficient values of the models constructed using the variables optimized by FDR and RFE algorithms on the point scale are improved by 0.933%~1.171%and 1.229%~1.542%respectively.Moreover,the phenomenon of spatial fragmentation of classification results is improved.Combining the verification accuracy at the point scale and the local spatial classification results,it was found that the FDR-RFE-RF model showed the most robust performance(OA=0.926,Kappa Coefficient=0.906),and the spatial distribution of impervious materials extracted for the whole study area was basically accurately represented the ground truth.From our results,we can conclude that variable select

关 键 词：Fisher可分比 递归特征消除算法 面向对象 不透水面材质 机器学习 航空遥感 

分 类 号：P237[天文地球—摄影测量与遥感] TP79[天文地球—测绘科学与技术]