基于GF-6卫星影像多特征优选的酿酒葡萄精准识别

Accurate recognition of wine grapes using multi-feature optimization based on GF-6 satellite images

多源遥感信息和特征优选是提高农作物识别精度的重要支撑,高分六号(GF-6)卫星作为首次引入红边波段的国产卫星,其丰富的光谱信息为作物识别提供了新的思路和解决途径。该研究基于宁夏回族自治区银川市永宁县2018年6月—2019年3月的GF-6数据,充分利用红边优势提取光谱特征、纹理特征和植被指数特征,构建多种特征组合方案,并根据随机森林算法对特征重要性进行度量,选取最优特征组合对酿酒葡萄进行精准识别。结果表明,与单一特征相比,多源遥感特征的增加显著改善了酿酒葡萄分类效果,其中,植被指数贡献程度最大,光谱特征次之;基于随机森林的优选特征组合分类效果最佳,其中,总体分类精度为94.15%,酿酒葡萄用户精度为94.23%,制图精度为92.59%;以实地调查的4个酒庄为验证区,将酿酒葡萄提取结果与统计数据进行对比,面积相对精度均在70%以上,其中优选特征结果相对精度在90%以上,研究结果将为国产卫星红边波段在植被分类和识别方面的应用提供数据参考。

Multi-source remote sensing information and feature optimization have become important supports to improve the accuracy of crop recognition.As the first Chinese satellite to introduce red-edge bands,the rich spectral information of GF-6 satellite provides new ideas for crop recognition.However,the use of crop features is confine to one single source,and previous studies on cash crop is relatively lacking.In this study,an available recognition method was proposed for the wine grape on multi-features using GF-6 satellite images.This paper first introduced the red-edge bands of GF-6 to the multi-source features in the study of accurate recognition for wine grape.Based on the GF-6 satellite data from June 2018 to March 2019 of Yongning County,Yinchuan City,Ningxia Hui Autonomous Region,two red-edge bands were selected to extract the spectral,texture,and temporal vegetation index features,including Normalized Difference Red-Edge Index(NDRE),Normalized Difference Vegetation Index(NDVI),and Enhanced Vegetation Index(EVI).A Random Forest algorithm with Gini index was used to choose the optimal number of features according to the measured importance scores,thereby to construct the optimal feature combination.Seven types samples of ground objects were selected to accurately recognize wine grape.Seven combinations of comparative features were designed,including three single-source and four multi-source feature combinations.Training samples and verification samples were obtained by the field investigation and visual interpretation with the Google Earth.The results showed that,compared with single-source features,the multi-source remote sensing features significantly improved the recognition effect of wine grape,where the vegetation index features contributed the most,then followed by the spectral and texture features.The accuracy of user and producer for the wine grape were 94.23%and 92.59%,respectively,the overall classification accuracy was 94.15%,and the Kappa coefficient was 0.93 in the optimal feature combination.Compared with spectral feature combination,the Jeffries-Matusita distance between wine grape-farmland and wine grape-woodland were improved from 1.57 to 1.99 and 1.40 to 1.97 respectively in the optimal feature combination..Compared with the combination 7,which includes all fifty-four features,the overall accuracy was improved by 2.85%in the optimal feature combination with only seventeen features.Taking four wine chateaus by field survey as the verification area,the results of wine grape extraction were compared with the statistical data,where the relative area accuracy of eight feature combinations were all above 70%,and that of the optimal feature combination was above 90%.Compared with other seven feature combinations,the optimal feature combination improved the separable measure of different ground objects and reduced field fragment,indicating more conformable with actual situation.In addition,the operation time of classification model was shortened,and the reasonable allocation of resources was realized by feature optimization.The successful launch of GF-6 enriched the existing satellite data sources,including red-edge bands,(such as RapidEye of Germany and Sentinel-2 of Europe).The findings can contribute to the large-scale remote sensing monitoring of wine grape and popularize the application of red-edge bands of Chinese satellite in agriculture,and also provide a sound reference to improve the performance of red-edge bands of Chinese satellite.