The Research Of Dimension Reduction Algorithm Based On Locally Linear Embedding And Its Applicaions In Precision Agriculture

China is a agricultural country with a population of billions, the problem of agriculture has always been one of the primary issue in the governments at all levels. Traditional agriculture in our country has been called the "intensive cultivation". This ensures our agriculture production have the advantages of higher per mu yield, but on the other hand, relying on artificial intensive cultivation purely must lead to low productivity problem. Modern agriculture is getting rid of the bondage of primitive agriculture、traditional agriculture and industrial agriculture, and entering into the knowledge agriculture development stage with the main characteristics of knowledge intensive.The Precision Agriculture which applied the modern information technology, biotechnology and engineering equipment technology applied in agricultural production has became the main production form of knowledge agriculture of every country in the new century.When the image processing and machine vision technology are applied in the precision agriculture, the intelligent analysis result of the images can be used to guide the robot to accomplish some field works automatically, which can improve the efficiency rapidly. But the information in optical image data is limited, it is not enough for many applications. Hyperspectral remote sensing image having plenty of bands, high spectral resolution and high pixel resolution, therefore it can provide more accurate information of ground objects, which has incomparable advantage over other datas.And in recent years, its application in precision agriculture has become increasingly widespread. For example, by satellite remote sensing technology hundreds of hectares of land are measured out the fertility of different plots, and control the agricultural machinery to complete the quantitative fertilization according to the local situation; the spectrum characteristics of crops can also be captured utilizing the ground remote sensing device. And these information can used to distinguish weeds from crops or judge degree damaged by diseases. It is hard to achieved relying on traditional agricultural methods.These new data analysis means has brought a revolutionary improvement to the agricultural production.But on the other hand,because of the huge datasize, not only the storage and transportation become a difficult task, but also the analysis and processing of datas have a greater challenges. So how to effectively reduce the dimension of data and the datasize is an important research subject in precision agriculture image analysis. This paper mainly studies the local linear embedding algorithm application to the problem of data dimension reduction in precision agriculture. Meeted the need of classification problem in the implementation of precision agriculture,such as weed identification, mainly around how to utilize the supervised information of learning samples in locally linear embedding algorithm, the adaptive selection of parameters, and the proper classification algorithm design were studied. The main research work and innovative results are as follows:(1)The basic theory of manifold learning method and developments are introduced.The influence of neighbor parametes、intrinsic dimension、noise and other issues to the dimension reduction effect is researched. The characteristics of the manifold method which is used commonly are analyzed, and the sensitivity to the parameters of them is compared. A kind of important research issue in the precision agriculture is to complete the automatic identification of some properties of the crops by intellectual technology, which is the typical applications of information technology and pattern recognition in precision agriculture. But the conventional locally linear embedding algorithm is an unsupervised algorithm, so its application in identify crop variety or diseases directly are often ineffective. A supervisied local linear embedding algorithm based on Fisher criterion is proposed. Firstly,the Fisher projection was carried out on the training samples to find out the best projection direction,and different kind of samples in this direction has the maximum separability. The projection distance of training samples in this direction is used to construct the neighborhood structure, which can make use of the training samples’ supervision information to instruct dimension reduction, so as to improve the recognition rate. The experimental results show that the supervied local linear embedding algorithm based on Fisher projection is more excellent than the conventional algorithm, so it can achieve high recognition rate only by some simple classification algorithm.(2)After the supervision problem is solved,there is another factor will affect the identification precision when the locally linear embedding algorithm is applied to the identification problems in precision agriculture, namely the neighbor parameter which is one of the main parameters in locally linear embedding algorithm. Whether the selection of this parameter is appropriate will seriously affect the recognition result. And this parameter selection is directly related to the characteristics of the datasets processed. There is no mature theory to direct this selection method currently, in most cases,the selection is obtained according to the result of many repeated experiments artificially. It has become a bottleneck in the development of local linear embedding algorithm. Aiming at the characteristics of data processed in the precision agriculture and the influence of neighborhood structure to recognition effect, the adaptive algorithm based on the supervised locally linear embedding. The experimental results show that this algorithm can ascertain neighbor parameter automatically according to the distribution characteristics of the dataset, on the premise of guarantee to obtain high recognition rate the algorithm efficiency is improved, so practicability is enhanced.(3)For classification problems, dimension reduction algorithm is just the first step, the another important link to ensure high recognition rate is the choice of classification algorithm. The locally linear embedding algorithm for the new test samples must repeat all steps again to finish dimension reduction with the training samples before classification, amount of calculation is large and the efficiency is low. Because the neighborhood structure is established according to the reconstruction error in the local linear embedding,a classification algorithm is used which compute the reconstruction error of the test samples versus the positive and negative manifolds and then judge the catigory of samples according to reconstruction error. This classification method is directly based on the characteristics of data’s manifold itself, and it does not introduce new unknown parameters, so it has the characteristics of easy application.(4)Weed identification is one of the main problems in application of precision agriculture. Because of the biological diversity in the nature, even if the same plants, there also has a certain differences on color and configuration, while different plants may be very similar. Using the traditional machine visual methods, by such as color and shape characteristics, the identification accuracy is not very high, and easily affected by the natural environment. Aimed at images aquired On corn field which have weeds and corn with complex symbiotic environment, a method is designed to segment weeds and corn automatically by the image morphology. Then using supervised locally linear embedding dimension reduction was carried out on the image after segmentation, the ideal experimental results were obtained. The local linear embedding algorithm based on Fisher projection also has the very good adaptability in the natural environment is proved.For the wheat blade hyperspectral datas which have rust disease collected in laboratory, according to the thought "the unity of the image and spectrum", a kind of image texture feature analysis method——gray symbiotic matrix(GLCM) is introduced, and conjoint analysis based on the GLCM and spectral information is carried out,so the advantages of imaging spectral datas are utilized fully. The experimental results show that this combination of traditional image analysis methods with the spectral method can recognize crops affected by the disease,especially in the early stage which can also be called recessive period, the identification effect is much better than that is obtained by the traditional spectral analysis method.