基于SVM与RF的无人机高光谱农作物精细分类

UAV Hyperspectral Remote Sensing Image Crop Fine Classification Based on SVM and RF

农作物精细分类是农业遥感的关键一步,无人机高光谱遥感为农作物精细分类提供了高效有利的途径.为了快速准确地实现农作物信息提取和分类,获取了研究区无人机高光谱影像,对影像开展一阶导数变换等预处理,并以此为基础对初始影像、一阶导数(FD)影像做MNF降维变换、特征波段选择等特征提取,从而获得初始影像、FD影像、MNF影像以及FD-MNF影像.最后应用SVM和RF建立了研究区4类典型农作物遥感判识模型,并采用Kappa系数和总体精度等评价参数对分析结果进行评价.研究结果表明,RF分类模型结果精度高达88%以上,同类影像RF分类结果精度比SVM高1%~5%,且与SVM方法相比RF对研究区农作物植株的提取效果更优;光谱曲线经一阶导数变换后只突出了大豆作物的光谱信息,导致分类结果精度降低;在所有分类模型中影像经MNF降维变换后分类效率及影像分类精度均有提高.采用无人机高光谱影像对研究区主要农作物进行精细分类,为后续研究区内农情监测等提供了有力的依据和支撑.

Fine classification of crop is a key step in agricultural remote sensing.UAV hyperspectral remote sensing provides an efficient and effective way for fine crop classification.In order to extract and classify crop information accurately.In this paper,we obtain the UAV hyperspectral images in the study area preprocess the images by using the first-order derivative transformation or other ways.Then the feature extraction of original image and first derivative(FD)image are carried out by MNF dimensionality reduction and feature band selection.According to the above steps,support vector machine(SVM)and random forests(RF)algorithm are used to build four types of classification models for crops:initial image,FD image,MNF image and FD-MNF image.The Kappa coefficient and the overall precision are used to evaluate the analysis results.The results show that the accuracy of RF classification results is over 88%.The accuracy of RF classification results of similar images is 1%-5%higher than SVM.In addition,compared with SVM method,RF has better extraction effect on crops in the study area.First-order derivative spectra only highlights the spectral information of the soybean crop,and causes the classification accuracy.MNF dimensionality reduction transform improves classification efficiency and image classification accuracy.Using UAV hyperspectral images to classify the main crops in the study area,provides a strong basis and support for subsequent monitoring of agricultural conditions in the study area.