基于Sentinel-2影像的喀斯特地区土地利用信息提取

Extraction of land use information in karst areas based on Sentinel-2 images

针对喀斯特地区分类尺度难以确定,特征数量维数过高,分类精度较低的问题,文章提出了通过联合评价确定最优分割尺度、ReliefF算法对先验特征数据集进行优选,使用分层掩膜的策略,利用随机森林算法完成分类的方法。并以贵阳西南喀斯特地区为研究区,首先使用同质性与Moran's I联合评价的方法确定最优分割尺度为80,通过ReliefF算法优选出重要性较高的15个特征;在此基础上,通过对比试验验证了随机森林算法的优越性;以2020年Sentinel-2影像为实验数据设计3种面向对象分类方案。结果表明,经最优尺度计算、特征优选和分层掩膜的分类方法结果精度最高,分类总体精度、Kappa系数、AD、QD分别达到0.886、0.849、0.092、0.022。最后将该方法应用于2023年影像,分类总体精度、Kappa系数、AD、QD分别达到0.868、0.825、0.106、0.026。证明了该方法在喀斯特地区土地利用信息提取方面的优越性和适用性。

Accurate land use information is the foundation of land management.Remote sensing data,characterized by its ease of acquisition,low cost,and high efficiency,has been widely used by scholars at home and abroad in the research of land use classification in combination with machine learning algorithms.Karst landforms are widely distributed in Southwest China.This region is fragile in ecological environment,due to its rugged terrain,large surface undulations,fragmented distribution of land plots.In addition,because of the long-term influence of its topography,the level of land use in the region is relatively low,and its economic development remains sluggish.Although extraction of accurate land use information is crucial for land resource management and planning in karst areas,the complex terrain and fragmented distribution of land plots in karst areas pose challenges for the extraction.Therefore,building on previous research,this study selected the southwestern part of Guiyang City,Guizhou Province-an area with karst landform characterized by complex terrain distribution and fragmented land plots-as a study area.With the use of Sentinel-2 satellite imagery as the basic data,the optimal object-oriented segmentation scale was calculated.The ReliefF algorithm was utilized to select features to input into the random forest algorithm,and land covers obtained from remote sensing images in different years were classified based on stratified classification.This study proposed a method that determined the optimal segmentation scale through joint evaluation,selected features from the prior feature dataset by random forest algorithm,and carried out classification with the use of ReliefF algorithm and a stratified masking strategy.Firstly,the optimal segmentation scale was determined as 80 by a joint evaluation with the combination of homogeneity and Moran's I.Subsequently,the ReliefF algorithm was employed to rank the importance of the initial features,with the top 15 significant features being selected.On this basis,the superiority of the random forest algorithm was verified by comparing multiple machine learning algorithms.Then,taking the Sentinel-2 images in 2022 as experimental data,this study designed three schemes for object-oriented classification to validate the superiority of the method with the combination of optimal segmentation scale,feature selection,and stratified masking in land use information extraction in karst areas.With the same samples,Model A completed the experiment by random forest algorithm without feature selection,and selected all 25 user-defined features,spectral features,shape features,and texture features.Model B selected the top 15 features after feature selection by ReliefF algorithm,and completed the random forest classification.Model C selected the top 15 features after ReliefF algorithm optimization and employed stratified masking under the random forest algorithm.In the principle of starting from the easiest area,non-vegetation areas were classified first.After the already classified areas had been ruled out by masking,the vegetation areas were classified,and finally,the classification results are merged.This method was then applied to the images in 2023 to verify its applicability in the extraction of land use information in karst areas.Through experimental research,the following conclusions can be drawn.Using the ReliefF algorithm to optimize the 25 classification features of remote sensing images can effectively improve classification accuracy and efficiency when training/validation samples are the same.In this study,the overall accuracy of Model B after feature selection was improved by 6.2%,compared to that of Model A with the original feature dataset,and Kappa was improved by 0.081.Multi-scale segmentation is the foundation of object-oriented classification and can avoid the"salt and pepper phenomenon".The evaluation of segmentation quality with the use of homogeneity and heterogeneity indices indicate that the optimal segmentation scale is 80.This approach can minimize the subjectivity of manual visual inspection.The random forest algorithm is superior in extracting land use information on different types of regions.Combining this algorithm with stratified masking can further reduce interference from already classified features to unclassified ones.The study demonstrates that the stratified masking method can achieve 88.6%of accuracy,the highest.