摘要
针对云南省祥云县林区云南松虫害区域高效识别的需求,为更加高效准确地对虫害信息进行监测,本文搭建了林区八旋翼多光谱图像采集平台,基于无人机多光谱图像提出了一种Jeffries-Matusita(J-M)距离优化的反向传播神经网络(BP)分类方法。该方法首先引入J-M距离实现了对训练样本的优化,有效降低了"同谱异物"和"同物异谱"现象的影响,然后基于颜色矩和灰度共生矩阵提取了图像的颜色和纹理特征,并提取了580、680、800 nm共3个波段的相对光谱反射率作为光谱值特征,建立了5个植被指数模型,最后利用BP神经网络算法对颜色、纹理、光谱值和植被指数4种特征向量进行训练识别,实现了对虫害区域的分类识别。利用所提算法从总体分类精度和Kappa指数两方面与传统BP神经网络和支持向量机(SVM)算法进行对比试验。试验结果表明,本文算法总体分类精度和Kappa指数分别达到了94.01%和0.92,建模时间相对于传统BP神经网络缩短了38%,总体分类效果优于传统BP神经网络和SVM算法。
In order to satisfy the needs of effective recognition in pest-affected region,a multispectral images acquisition platform was built to monitor the pest-related information efficiently and accurately in Yunnan pine forest region of Yunnan Province. Aneural network of Jeffries-Matusita（ J-M） distance optimized back-propagation（ BP） neural network was proposed based on unmanned aerial vehicle（ UAV）multispectral images. Firstly,the method realized the optimization process of the training samples by introducing the J-M distance concept,which reduced the influence of both ＂similar spectral from multiple objects＂and ＂multiple spectral from similar objects＂. Then,the color and texture features of the images were extracted based on their color and the gray-scale co-occurrence matrix. Three bands of relative spectral reflectance,namely 580 nm,680 nm and 800 nm were extracted as spectral characteristics.Meantime,five vegetation index models were established to identify pest area. Finally,BP neural network algorithm was applied for training and identifying four feature vector quantities,including color,texture,spectral and vegetation index,which greatly achieved the identification and classification goal of pest region. The proposed algorithm was compared with the traditional BP neural network and support vector machine（ SVM） algorithm from both general classification precision and the Kappa index. The experimental results showed that the overall accuracy index of classification and the Kappa index of the algorithm reached 94. 01% and 0. 92,respectively,which was superior to traditional BP neural network and SVM algorithm. Besides the modeling time was shortened by 38% when compared with the traditional BP neural network method. The improved efficiency satisfied the high efficiency identification needs of Yunnan pine pest area in Xiangyun County.
作者
张军国
韩欢庆
胡春鹤
骆有庆
ZHANG Junguo1,HAN Huanqing1,HU Chunhe1,LUO Youqing2(1. School of Technology, Beijing Forestry University, Beijing 100083, China 2. College of Forestry, Belting Forestry University, Belting 100083, Chin)
出处
《农业机械学报》
EI
CAS
CSCD
北大核心
2018年第5期249-255,共7页
Transactions of the Chinese Society for Agricultural Machinery
基金
北京市科技计划项目(Z171100001417005)
中央高校基本科研业务费专项资金项目(2016ZCQ08)
