基于改进SVM算法的典型作物分类方法研究

Study on classification method of typical crops based on improved SVM algorithm

以黑龙江省哈尔滨市阿城区为研究区域,多时相16 m空间分辨率高分一号(GF-1)卫星宽覆盖(Wide field of view,WFV)影像为数据源,选择归一化植被指数(Normalized difference vegetation index,NDVI)、增强植被指数(Enhanced vegetation index,EVI)、归一化水指数(Normalized difference water index,NDWI)、比值植被指数(Ratio Vegetation Index,RVI)4种植被指数,构建植被指数时间序列,分析作物特征曲线,结合实地样本数据,采用支持向量机(Support vector machine,SVM)分类器对研究区内主要农作物玉米、水稻和大蒜/白菜实施分类。针对SVM分类器分类精度较低问题,引入自适应变异粒子群算法(Adaptive mutation particle swarm optimization,AMPSO)优化SVM,克服传统SVM参数选择主观性,进而提升分类器分类精度。结果表明,玉米和水稻生育期与大蒜/白菜差异较大,易区分;玉米与水稻生育期接近,光谱信息相似,区分难度较大,但光谱指数增长与回落趋势不同,借助NDVI、RVI和EVI可实现有效区分。改进后的AMPSO-SVM分类器,分类效果相比于SVM明显提升,确定核参数为0.135,惩罚因子为221.67时,分类效果最佳,总体分类精度达到94.39%,Kappa系数为0.9287,比SVM分类器,分类精度提升3.48%,Kappa系数提高0.0436。研究可为大区域农作物种植结构提取提供参考与借鉴。

Taking Acheng District,Harbin,Heilongjiang Province as the study area,data source came from the wide coverage images of GF-1 satellite with multi-temporal 16m spatial resolution.Four vegetation indexes were applied in this paper,namely,Normalized Difference Vegetation Index,Enhanced Vegetation Index,Normalized Difference Water Index and Ratio Vegetation Index,to construct vegetation index time series,crop characteristic curve and combine field sample data were analyzed.Support Vector Machine was applied to classify the main crops corn,rice and garlic/cabbage in the study area.In view of the low classification accuracy of SVM classifier,the adaptive mutation particle swarm optimization algorithm was introduced to optimize SVM.It effectively overcame the subjectivity of traditional SVM parameter selection and then improved the classification accuracy of the classifier.The results showed that the growth period of garlic/cabbage was significantly different from that of corn and rice,and it was easy to distinguish.The growth period of corn was close to that of rice,and their spectral information was similar,so it was difficult to distinguish between them.However,the growth of the spectral index was different from the downward trend.With the help of NDVI,RVI and EVI,an effective distinction could be achieved.In addition,the improved AMPSO-SVM classifier had a significantly improved classification effect compared with SVM.When the kernel parameter was 0.135 and the penalty factor was 221.67,the classification effect was the best desired.The overall classification accuracy reached 94.39%,and the Kappa coefficient was 0.9287,which was 3.48%higher than the SVM classifier,and the Kappa coefficient was increased by 0.0436.Therefore,it is hoped that this method can provide reference for the extraction of crop planting structure in large areas.