基于多分类器集成的GF-2作物信息精细提取研究

Refined extraction of GF-2 crop information based on multi-classifier integration

【目的】将多种分类器的优点融合,以便提升遥感影像作物信息提取的精度。【方法】以渭库绿洲为研究区,利用国产高分2号(GF-2)数据和野外调查数据,基于提取的遥感识别特征制定不同分类方案,采用马氏距离(Ms DC)、最小距离(MDC)、最大似然(MLC)、神经网络(NNC)、支持向量机(SVM)5种传统机器学习方法分别对6种特征组合方案的影像进行分类,然后选择基分类器,并应用多数投票法和保守投票法2种多分类器集成算法,对研究区农作物进行精细分类提取。【结果】①辅助特征的加入对于子分类器的精度提高明显。5种分类器中除了MLC,其余4种分类器都是在加入归一化植被指数特征(NDVI)和纹理特征后取得了最高精度。②基分类器中精度最高的是NNC-4(人工神经网络的第4种特征组合方案),OA达到83.54%,Kappa系数为0.77。③相比基分类器,多分类器集成方法能够在制图精度和用户精度两方面提高农作物的提取精度。并且保守投票法优于多数投票法,OA为85.89%,Kappa系数为0.80。④集成分类结果中除了棉花的识别精度与最优基分类器NNC-4相等,达到94.94%外,其他的农作物如套种棉花、玉米、套种玉米、核桃园的识别精度都高于NNC-4,其中套种玉米与核桃园的提取效果较好,精度分别达到86.05%、79.09%;对于套种棉花的提取较差,只有63.86%;玉米的提取最差,只有12.17%。【结论】本文应用GF-2数据,基于多分类器集成方法对复杂背景下的多种作物及种植结构进行精细提取研究,拓展了作物信息提取的方向和GF-2数据的应用领域。

【Objective】The advantages of several classifiers were combined to improve the accuracy of crop information extraction from remote sensing images.【Method】The paper took Weiku Oasis as the research area,used the domestic GF-2 data and field survey data,and formula-ted different classification schemes based on the extracted remote sensing identification features.Five traditional machine learning methods as Mahalanobis distance(Ms DC),minimum distance(MDC),maximum likelihood(MLC),neural network(NNC),and support vector ma-chine(SVM)were used to classify the images of the six feature combination schemes.Then the base classifier was selected,and two multi-classifier integration algorithms,the majority voting method and the conservative voting method were applied to finely classify and extract the crops in the study area.【Result】(i)The additional auxiliary features significantly improved the accuracy of sub-classifiers.Among the five classifiers,except MLC,the other four classifiers achieved the highest accuracy after adding NDVI and texture features.(ii)The most accu-rate base classifier was NNC-4(the fourth feature combination scheme of artificial neural network),with OA reaching 83.54%and Kappa coefficient of 0.77.(iii)Compared with the base classifiers,the multi-classifier ensemble method could improve the extraction accuracy of crops in both producer’s accuracy and user’s accuracy.And the conservative voting method was better than the majority voting method,with OA being 85.89%and Kappa coefficient being 0.80.(iv)In the integrated classification results,except that the recognition accuracy of cot-ton was equal to the optimal base classifier NNC-4,reaching 94.94%,the recognition accuracy of other crops such as interplanted cotton,corn,interplanted corn and walnut orchards was higher than that of NNC-4.Among them,the extraction accuracy of interplanted corn and walnut orchard was better,with the accuracy reaching 86.05%and 79.09%respectively;The extraction of interplanted cotton was poor,only 63.86%;The extraction of corn was the worst,only 12.17%.【Conclusion】In the study,the GF-2 data was applied,and the multi-classifier ensemble method was used to conduct fine extraction research on various crops and planting structures under complex backgrounds,which expanded the direction of crop information extraction and the application field of GF-2 data.