基于CNN和SVM的无人机多光谱遥感草地植物识别

为选择最佳的物候期、飞行高度和识别模型提高植物识别的精度，本研究以伊犁绢蒿(Seriphidium transiliense)荒漠草地主要植物伊犁绢蒿、角果藜(Ceratocarpus arenarius)以及裸地为识别对象，选择4月、6月、9月3个飞行时期，15m，30m，60m 3个飞行高度，通过无人机搭载多光谱相机采集草地群落多光谱数据，在分析光谱反射率差异的基础上，利用最佳指数因子(Optimum index factor，OIF)筛选特征波段，通过卷积神经网络(Convolutional neural network,CNN)和支持向量机(Support vector machines,SVM)建立识别模型。结果表明：地物反射率4月＞6月＞9月，15m＞30m＞60m；不同飞行高度下OIF值一致，但在月份间具有差异，4月敏感波段为Green，Red和NIR，6月和9月敏感波段为Red，Red edge，NIR；在识别精度上SVM＞CNN，4月＞9月＞6月，15m＞30m＞60m，裸地＞伊犁绢蒿＞角果藜。综合来看，采用SVM在4月、15m飞行高度下进行识别的总体精度最高，达到86.23%。

Identification of Grassland Plants by UAV Multispectral Remote Sensing Based on CNN and SVM

To improve the accuracy of plant identification by selecting the best phenological period,flight height and identification model,the main plants Seriphidium transiliense, Ceratocarpus arenarius and bare land on Seriphidium transiliense desert grassland were taken as the identification objects in this study. Through three flight periods in April,June and September and three flight altitudes of 15 m,30 m and 60 m,the multispectral data of grassland community through the multispectral camera carried by UAV were collected. On the basis of analyzing the difference of spectral reflectance,the characteristic bands were screened using optimum index factor value,and the identification model was established through convolutional neural network and support vector machines. The results showed that the orders of surface reflectance in different months and at different heights were April>June>September and 15 m＞30 m＞60 m,respectively. OIF values at different flight altitudes were consistent,but different between months. The sensitive bands were green,red and NIR in April,and red,red edge and NIR in June and September;In terms of recognition accuracy,the orders under different conditions were SVM>CNN,April>September>June,15 m>30 m>60 m,and bare land>Seriphidium transiliense>Ceratocarpus arenarius. Overall,the accuracy of identification using SVM was the highest at the flight height of 15 m in April,reaching 86.23%.