基于CNN和SVM的地面高光谱遥感草地植物识别  

Identification of grassland plants using hyperspectral remote sensing based on convolutional neural network and support vector machine

作　　者：马建 刘文昊 靳瑰丽[1] 宫珂 刘智彪 李莹 李嘉欣 王生菊 MA Jian;LIU Wenhao;JIN Guili;GONG Ke;LIU Zhibiao;LI Ying;LI Jiaxin;WANG Shengju(College of Grassland Science,Xinjiang Agricultural University/Xinjiang Key Laboratory of Grassland Resources and Ecology/Key Laboratory of Grassland Resources and Ecology of Western Arid Region,Ministry of Education,Urumqi 830052,Xinjiang,China)

机构地区：[1]新疆农业大学草业学院/新疆草地资源与生态重点实验室/西部干旱荒漠区草地资源与生态教育部重点实验室,新疆乌鲁木齐830052

出　　处：《草业科学》2023年第2期394-404,共11页Pratacultural Science

基　　金：国家自然科学基金项目(31960360)。

摘　　要：物候期和识别模型的选择直接影响植物识别的精度。本研究以蒿类荒漠草地主要植物伊犁绢蒿(Seriphidium transiliense)、角果藜(Ceratocarpus arenarius)以及裸地为识别对象,选择4月、6月、9月3个时期,通过SOC 710 VP高光谱成像仪采集草地群落高光谱数据,在分析地物光谱反射率差异的基础上,利用最佳指数因子(OIF)筛选特征波段,通过卷积神经网络(CNN)和支持向量机(SVM)建立识别模型。结果表明:1)不同物候期的伊犁绢蒿与角果藜在可见光波段均表现为“低-高-低”的光谱反射率趋势,并随月份增加峰谷现象逐渐不明显;红边波段这两种植物表现出快速上升;在NIR平台区4月各识别对象间反射率大小差异最明显。2)利用OIF筛选的识别波段组合在月份间表现一致,为638.64、789.49和923.79 nm。3)在识别精度上,SVM>CNN;4月>9月>6月;裸地>伊犁绢蒿>角果藜。综合来看,采用SVM在4月对蒿类荒漠草地主要植物进行识别的精度最高,为92.12%。Selection of the phenological period and identification model directly affects the accuracy of plant identification.In this study,the dominant plants Seriphidium transiliense and Ceratocarpus arenarius in sagebrush desert grassland and bare land were used as identification objects.We collected grassland community hyperspectral data in April,June,and September using an SOC 710 VP hyperspectral imager to analyze the differences in ground object spectral reflectance.The optimum index factor was used to screen feature bands,and an identification model was established using convolutional neural network(CNN)and support vector machine(SVM).The results showed that:1)In the visible light band,the spectral reflectance of the two species showed a“low-high-low”trend,and the peaks and valleys became less obvious as the month progressed.In the Red Edge band,species 2 increased rapidly.The near-infrared platform area in April showed most obvious difference in reflectance between the identified objects.2)The bands selected using the optimum index factors were at 638.64,789.49,and 923.79 nm.3)The order of identification accuracy was as follows:SVM>CNN,April>September>June,bare land>S.transiliense>C.arenarius.Overall,SVM was the most accurate method for identifying the dominant plants in sagebrush desert grassland in April,with an accuracy of 92.12%.

关 键 词：高光谱 伊犁绢蒿 角果藜 裸地 特征筛选 识别方法 深度学习 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置] TP18[自动化与计算机技术—控制科学与工程] S812[农业科学—草业科学]