基于SVM的地面成像光谱数据田间杂草识别研究

SVM-based Weed Identification Using Field Imaging Spectral Data

为提高当前田间杂草识别精度,利用地面成像光谱数据研究多特征参与的SVM田间杂草识别方法,根据地面成像光谱数据的特点提取田间作物杂草的多种可区分特征,包括纹理特征、连续统去除后的光谱特征和高光谱植被指数特征,并对高维特征集进行降维,利用多种多特征组合参与SVM田间杂草识别。实验结果证明,在训练样本一致的前提下,多特征参与的SVM田间杂草识别精度优于仅使用原始光谱特征时的情况,使用包含原始光谱特征、纹理特征、高光谱植被指数特征和连续统去除后的光谱特征的多特征组合时田间杂草识别精度最高。

Weed identification is a fundamental task for precision agriculture as well as a premise for variable spraying and accurate weeding.With both of high spectral resolution and high spatial resolution,the use of the field imaging spectral data on weed identification has a bright future.Compared with any other classification and identification method,the SVM algorithm enjoys many advantages.Since the uncertainty of the object spectrum,the use of the original spectral features only may restrict the accuracy of the SVM classification and identification.In order to improve the accuracy of the weed identification,multiple features were involved in the SVM-based weed identification using imaging spectral data.Features that enable weed identification were extracted.Meanwhile,the feature dimension was reduced.This experiment shows that the SVM-based weed identification with multiple features is more accurate than the SVM-based weed identification with original features only.