基于面向对象结合随机森林模型的Sentinel-2A影像耕地信息提取

Sentinel-2A image cultivated land information extraction based on object-oriented and random forest model

为了更快、更准确地提取耕地信息,以山东省青岛市莱西市夏格庄镇为研究区,利用Sen⁃tinel-2A影像融合光谱特征、遥感指数特征、纹理特征和形状特征等31个特征变量,设计4种耕地信息提取方案,采用结合面向对象的随机森林(Random Forest,RF)分类模型提取耕地信息,并基于相同分类条件,与传统机器学习分类方法对比,评价模型的优适性。结果表明:结合所有分类特征变量的方案4耕地提取效果最佳,其中旱地提取精度高达99.6%,大棚提取精度达88.4%;5种分类方法中,结合面向对象的RF模型耕地提取精度最高,减弱了分类结果的“椒盐”现象,优化了分类结果。

Taking Xiagezhuang Town,Laixi City,Qingdao City,Shandong Province as the study area,using Sentinel-2A imagery as the data source,integrating 31 feature variables including spectral features,remote sensing index features,texture features and shape features,and designing four types of cultivated land infor⁃mation extraction in the scheme,an object-oriented random forest(Random Forest,RF)classification model was used to extract farmland information.Based on the same classification conditions,a comparative experi⁃ment with traditional machine learning classification methods was carried out to evaluate the suitability of the model.The results showed that:Scheme 4,which combined all the classification feature variables,had the best cultivated land extraction effect.The dryland extraction accuracy was as high as 99.6%,and the greenhouse extraction accuracy was 88.4%.Among the five classification methods,the object-oriented RF model had the highest extraction accuracy for cultivated land.The"salt and pepper"phenomenon of the classification result was weakened,and the classification result was optimized.