遥感影像分类是森林资源调查和监测不可缺少的内容，分类的精度直接影响遥感数据的应用水平和实用价值。随着遥感技术的发展，卫星上搭载的传感器已能获得高分辨率的数字影像。在高分辨率卫星影像中，目标地物的细节比TM和MSS影像更为清晰，呈现出大量的纹理与结构信息。如何解决多类别图像识别并满足一定的精度，是高分辨率遥感图像研究中的一个关键问题，具有十分重要的意义。本文在国内外森林植被遥感分类研究的基础上，应用快鸟影像，用径向基神经网络和BP神经网络方法对广西地区选取的样地进行地类分析研究，提取研究区域的林地像元，并与传统的模式识别方法（最小距离法）的提取结果进行精度比较。主要研究工作和成果如下： （1）本文以快鸟图像作为主要数据源，将径向基神经网络和BP神经网络用于森林植被像元的提取，提取精度分别达到了95.13%和96.26%，表明对林地像元的分类质量很好。 （2）本文针对水田、农田等对林地像元提取影响较大的样地类型进行了研究，通过分析水田和农田样地像元灰度值的分布特征，找出其与林地像元灰度值分布的差别之处。研究发现，以相邻三个像元灰度值相同作为水田和农田样地像元的判别标准，可以很好的将其与林地像元区别开。通过对水田样地的验证，这种方法完全可行。 （3）应用设计函数 设计RBF神经网络，对12个林地样地进行网络训练，确定径向基神经网络散布常数的最佳值为0.5。设计径向基神经网络，对分类样地进行林地像元提取，提取精度达到了95.13%，表明分类质量很好。 （4）应用不同的训练函数对BP网络进行训练，比较训练误差和时间，最终选取训练函数（trainrp）为BP网络的训练函数；对12个林地样地进行网络训练，确定BP网络的神经元数为25；设计BP神经网络对分类样地进行林地像元提取，提取精度达到了96.26%。

The classification on the remote sensing image is the indispensable content to the forest resource investigation and supervisal. The classification precision influences application level and practicality value of the remote sensing data directly. With the development of remote sensing technology, the sensors used on planets can gain high resolution digital image. In the high resolution remote sensing image, the details of objections are more clear than TM and MSS images, showing great texture and structure information. The key question in the research on high resolution remote sensing image is how to identify the image from many kinds of objections and meet stated precision, which has very important meaning. Base on the classification research of the forest vegetation both here and abroad, the text uses radial basic function neural network method and BP neural network method to analysis and study the types of selected sample on the area of Guangxi province quick bird image. The text picks up the forest pixels in research area using Neural Network method, and compares the precision with conventional mode identify method (least distance method). The main research job and result as follows: (1)This text uses radial basic function neural network method and BP neural network method to pick up the forest pixels on the quick bird image, and the precision achieves 95.13% and 96.26% each other. The result shows the classification quality is satisfaction using the two methods. (2)The text does main study to some sample types, which have great influence to the pick-up forest pixels such as paddy field, farmland and so on. The text researches the distributing character of the pixels gray value in the paddy field and farmland, and finds out the difference with forest pixels gray value distribution. Making the gray value equal of three border upon as the criterion to identify the paddy field and farmland pixels, we can differentiate those pixels from the forest pixels. Tested to the paddy field, the method is feasible. (3) The text uses the design function to design RBF neural network, and trains the RBF neural network using 12 forest samples. The study ascertains that the best spread value of RBF neural network is 0.5. Then we use this RBF neural network to pick up the forest pixels in the classification sample, and the precision reaches 95.13%. (4) The text uses different training function to train the BP neural network, compares training error and training time, and selects training function (trainrp) as the training function of BP neural network. Training the BP neural network using 12 forest samples, we ascertain the count of nerve cell is 25. Then we use this BPNN to pick up the forest pixels in the classification sample, and the precision reaches 96.26%.