Research On Weeds Classification Based On Hyperspectral Image

With the development of observing technology about surface features, remote sensing technology has widely applied to various of fields. Especially in agriculture. Classification in images is the core part in high spectrum., and how to classify surface features quickly and rightly is the key point. In traditional classification model, the requirement of sample is inclined to infinity point, which is the best environment that the machine realize classification function. However, it is hard to reach this requirement in reality. Especially hyper-spectral image has following challenges in classification technique, including huge amount of wave band, high figure of dimension, large quantity of data, small size of sample. For the pursuit of best result about classification, this essay is doing classification research on hyper-spectral image, based on classification in support vector machine. Main contents included:Firstly, research background is introduced, and so does current status and promotion in kinds of methods, which are common methods to classify hyper-spectral image. Then data features about hyper-spectrum is analyzed from the aspect of its basic theory and data. Secondly, make a brief introduction about the classification process, and point out possible problems. Which provide theoretical foundation to further research.Secondly, the process of image classification, per-treatment of dimension reduction about image are introduced. And as nonlinear structure, there is utilizing IISOMAP-LLE manifold algorithm. Through experiment, this method, based on feasibility, can result in better nonlinear structure, which is hiding in weeds. Finally, do an analysis of commonly unsupervised classification and supervised classification, so as to compare performance about each classification method.Finally, introduce the weeds’ image classification based on Supporting Vector Machine, it is concluded that Supporting Vector Machine has better characteristic of classification in small sample, non-linearity, high dimension. Kernel function is the core of Supporting Vector Machine. And the normal kernel functions consist of Linear kernel, Polynomial kernel, Radial Basis Function kernel and Sigmoid kernel. Based on the analyze of them, kernel function also can be divided into Global Function and Local Function。Different kernel functions were reorganized as a new kernel function. Then the combined kernel function and normal kernel function were applied to the Supporting Vector Machine. Compared to experiment, it is the best conclusion to combine Polynomial kernel with the global feature and Radial Basis Function kernel with the Local features, which has remarkably promoted in Overall Accuracy and Kappa Coefficient. The new combined kernel function, with better capacity for learning, which fully utilizes the special information from hyperspectral image, is helpful for hyperspectral image to apply to agriculture.