基于GF-1/WFV时间序列的葡萄识别模型——以宁夏红寺堡区为例

Grape Recognition Model Based on GF-1/WFV Time Series:A Case Study in Hongsibu District of Ningxia

以宁夏红寺堡区为研究区,基于高分一号(GF-1/WFV)卫星构建葡萄生长季时间序列光谱数据,运用(Jeffreys-Matusita)(J-M)距离分析葡萄地块归一化植被指数(NDVI)时序曲线特征确定了最佳识别时相,将最佳时相的NDVI、相邻时相差值速率和曲线积分训练样本集导入Clementine数据挖掘软件中,利用C5. 0决策树分类算法,并结合专家经验法构建葡萄林决策树提取模型。结果表明:构建的识别模型能够满足葡萄的识别需求,但在不同覆盖度的葡萄地块上精度有所差异;基于决策树分类的总体精度为93. 71%,Kappa系数为0. 91。其中,中低覆盖度葡萄林生产精度为90. 82%,用户精度为88. 56%;高覆盖度葡萄林生产精度为92. 44%,用户精度为91. 18%。

Grape is one of the most widely distributed fruit tree species,and its accurate spatial distribution is of great significance for the management and development of grape planting and wine industry.In this study,the Hongsibao in Ningxia was taken as the study area to obtain the time series of spectrum data in grape growing season based on the Gaofen-1 satellite(GF-1/WFV),the J-M distance analysis was used to analyze the normalized difference vegetation index(NDVI)of the grape plots,and the best recognition phase was determined.The NDVI values of the best phase,adjacent to the value difference rate and curvilinear integral training samples were put into the Clementine data mining software.The extract grape model of decision tree was developed by using C5.0 decision tree classification algorithm and combining with the expert experience method.The results suggested that the established recognition model could meet the needs of grape recognition,but the accuracy was different from different grape plots due to their different coverage.The overall accuracy based on the decision tree classification was 93.71%,and the Kappa coefficient was 0.91.In which the production precision and users’precision of the grape plots with moderate or low coverage were 90.82%and 88.56%,and those with high coverage were 92.44%and 91.18%,respectively.