基于光谱和神经网络模型的作物与杂草识别方法研究  

Identification Methods of Crop and Weeds Based on Vis/NIR Spectroscopy and RBF-NN Model

作　　者：朱登胜[1] 潘家志[2] 何勇[2] 

机构地区：[1]金华职业技术学院信息工程学院,浙江金华321017 [2]浙江大学生物系统工程与食品科学学院,浙江杭州310029

出　　处：《光谱学与光谱分析》2008年第5期1102-1106,共5页Spectroscopy and Spectral Analysis

基　　金：国家科技支撑计划项目(2006BAD10A09);国家自然科学基金项目(30671213);浙江省重大科技攻关项目(2005C12029);浙江省三农五方资助

摘　　要：利用光谱技术来识别作物与杂草是精细农业中一个非常重要的研究内容，但光谱数据中含有大量冗余数据，如何预处理以及建立识别模型，是决定识别准确率的关键。利用在325-1075nm波段的光谱识别了三种杂草（牛筋草、凹头苋、空心莲子草）与大豆幼苗。在幼苗生长的第三周与第六周分别采集杂草与作物的光谱，共378个样本。用其中的250个光谱样本，包括第一期和第二期采集的光谱样本，在采用db12小波经过三层分解后，将其小波系数作为输入数据建模，构造了一个径向基函数神经网络。然后，利用余下的光谱样本检验该模型的识别能力。结果表明，该模型对作物与杂草光谱具有极强的识别能力，只有3个第二期的牛筋草样本被判断为空心莲子草，其余的样本全部正确识别。这个结果表明，采用可见／近红外光谱识别大豆幼苗与三种伴随生长的杂草是可行的，同时也说明，随着作物的生长阶段的不同，其光谱的变化不会影响到种类识别。The automated recognition of crop and weed by using Vis/NIR spectral in field is one of hottest research branches of agriculture engineering. If the recognition is efficient and effective, then the variate operations of herbicide or fertilizer spraying in field could be realized. Many researches have pointed out that the reflectance rate of green plant leaves could be used to identi- fy the varieties. As the colors and surface textures of crop and weed were change in different living phases, these changes may exert great influence on the reflectance spectral of plant leaves. Vis/NIR spectra of three weeds and one crop in two different terms were recorded by spectral meter ASD FieldSpec Pro FtL Its wave band is from 325 to 1 075 nm. The scan time was 270 ms. The scanning times of per sample was set to 30 times. Firstly, 23 days after the planting of soybean, some soybean leaves and weeds leave were picked from the field, and brought to lab to record spectral. The lighting condition was controlled by an ar- tificial halogen bulb. Secondly, on the 45th day, the same experiment was done. The three weeds were goosegrass, alligator al- ternanthera and emarginate amaranth. The crop was soybean seedling. Totally 378 samples were drawn for two terms. As one original reflectance spectrum contains 651 float numbers, the total data volume was huge. Using wavelet transform to compress data volume and extract characteristic spectral data was tried. The result was 114 float numbers per sample. Among them, 250 samples from two terms were used as input data to build artifieial neural network model, and those left were used to cheek the validation. Radial basis funetion neural network model is widely used in pattern recognition problems. It is a nonlinear and self a- daptive parallel. By assigning a 1 by 4 veetor to eaeh speetral samples, the souree data could be used to build an RBF-NN model. All the samples were assigned these standard output data. Then, the left 128 samples were used to eheek the performanee of the model. The r

关 键 词：近红外光谱 RBF人工神经网络 识别 杂草 豆苗 

分 类 号：S52[农业科学—作物学] TH74[机械工程—光学工程]