随机森林方法在玉米-大豆精细识别中的应用

研究基于遥感影像的作物精确识别技术方法,对获取作物分布信息具有重要意义。随机森林分类(random forest classification,RFC)是机器学习的一种,本文使用Landsat-8 OLI卫星影像数据,针对研究区内的大豆、玉米和其他地物等3种主要作物类型,系统比较了该方法与较为成熟的最大似然分类(maximum likelihood classification,MLC)、支持向量机分类(support vector machine,SVM)方法的分类精度。结果表明,MLC、SVM、RFC的总体分类精度分别为91.68%、91.49%、94.32%,Kappa系数分别为0.87、0.87、0.91,RFC方法作物识别精度比MLC和SVM分类显著提升。对原始7波段影像进行主成分变换(principal component analysis,PCA),提取前4个主成分分量,同时计算归一化植被指数(normalized difference vegetation index,NDVI)和归一化水体指数(normalized difference water index,NDWI),将6个额外辅助特征波段叠加到原始7个波段影像上进行再次分类,MLC和SVM方法作物识别精度未有提升,RFC方法总体精度提高了1.49个百分点,Kappa系数提高0.03,精度提升幅度有限,主要原因是6个辅助波段在类型识别中作用较小。在分类耗时上,MLC、SVM、RFC分别为145 s、11 000 s、1800 s,表明随机森林分类具有最好的分类精度和适中的耗时。综合评价后,随机森林分类方法在进行大豆-玉米精细识别中具有较大优势,具有业务应用的潜力。

Application of Random Forest Method in Maize-soybean Accurate Identification

It is very important to obtain the crop identification information based on remote sensing image. Remote sensing images have the advantages of high efficiency, high accuracy, low costs, and wide monitoring scope. Applying remote sensing images in maize-soybean accurate identification and planting area evaluation can give full play to the advantages of remote sensing images. Random forest classification (RFC) is a new classification method, a type of machine learning. Currently, there are very few studies on crop classification based on RFC. In order to evaluate the potential of the method on maize-soybean crop accurate identification, the paper conducted classification of major crops of soybean, maize, and other ground objects. Utilizing Landsat-8 OLI satellite image data, and three methods including maximum likelihood classification (MLC), support vector machine (SVM), and random forest classification (RFC). The overall classification accuracies of MLC, SVM, and RFC were 91.68%, 91.49%, and 94.32%, with their kappa coefficients of 0.87, 0.87, and 0.91, respectively, showing that RFC is better. The principal component analysis (PCA) was made on original seven wave band images, and the first four wave bands of the major components were extracted. Meanwhile, the normalized difference vegetation index (NDVI) and normalized difference water index (NDWI) were calculated; six additional supporting characteristic wave bands were overlapped on original seven wave band images, and the classifications with MLC, SVM, and RFC were conducted again. After adding characteristic wave bands, crop identification accuracies by MLC and SVM methods were not improved. The accuracy of RFC method was increased slightly with overall accuracy of 95.81% increasing by 1.49 percent, and Kappa coefficient of 0.94 increasing by 0.03, showing accuracy slightly increased, and limited improvement effect. Near-infrared band and two short infrared wave bands were most important, while newly added wave band was not significant for soybean-maize identification, showing the limited improvement effect of supporting wave band. SVM had the longest time spent on classification, with about 11 000 s; MLC the least, only 145 s; and RFC about 1800 s. It indicates that SVM doesn’t have any advantages in both accuracy and time-consumed, however, MLC can quickly get the classification results, and RFC has the highest classification accuracy with moderate time consumed. In conclusion, RFC has greater advantage in soybean-maize accurate identification, and is suitable to be widely applied in the operation of regional agriculture remote sensing monitoring crop area extraction.