选取6种输入向量组合，利用线性网络、BP网络以及径向基网络等3种神经网络模型进行比较研究，筛选最适宜网络模型和最佳输入组合，建立叶片数字图像彩色信息和叶片氮含量的关系模型，探索利用神经网络技术获取叶片数字图像信息的方法。结果表明，径向基网络在利用数字图像(B，H，G-R，G/R)指标作为网络输入向量时，能够实现获取棉花叶片数字图像氮含量的目标。径向基网络训练的180组样本的训练精度均达到极显著水平（r = 0.9022**），30组测试样本的预测值与实测值也达到极显著相关（r = 0.8674**），径向基网络和(B，H，G-R，G/R)向量是一种适合本研究的数学模型。对利用神经网络提取棉花叶片数字图像氮含量技术的初步探索，拓展了神经网络和数字图像技术在农业生产中的应用。
Artificial Neutral Network (ANN) has some important features, such as self-study, acceptance-error, building math model rapidly. ANN has been widely used in many fields, some people have made a lot of findings in agriculture by ANN. The technology of digital image processing is also very important for agriculture, and people have found there are some relation between color information and the nitrogen quantity for maize, tomato. But nobody use ANN to found the relation. The objective of this research is to process the digital image of cotton leaf, and use ANN to select the best math model and input vectors for establishing the relation between the color information and nitrogen quantity of cotton leaf. So we can use the advantages of ANN and the technology of digital image processing, and select the most suitable result for this research automatically. We select three ANN models (line on network, BP network and radical basis function (RBF) network) and six pieces of input vectors for this research, and train each model with color information from 180 pieces of digital images, and use the better to forecast nitrogen quantity of 30 pieces of images. The results showed that linear network was not fit this research and the relation between color information and nitrogen quantity was not fit the linear models, and RBF network was better for this research than BP network. RBF network had a lot of advantages in calculating the quantity of nitrogen using vector (B, H, G-R, G/R). The precision of training result was very marked, with r = 0.9022**, and the precision of forecast was high, with r = 0.8674** by this ANN forecast using the 30 pieces of cotton digital image. Because of local smallest, simple framework, and rapid training, RBF network can get the nitrogen quantity in plant by digital image information, and enhance the application of ANN in agriculture.
全 文 ：