Discriminant Analysis of Hyperspectral Data for Bio Species Recognition

高光谱数据包括来自地面或机载光谱仪的数百个波段的强度读数。本文拟利用(1)野外直接从森林冠层上方测量的高光谱数据，以及(2)基于第(1)部分理解的机载仪器高光谱数据，对森林物种进行识别。本研究的动机是高光谱数据中包含的丰富的光谱信息可以提高森林物种的识别水平，希望找到一种相对简单的方法，既能处理高光谱数据中大量的光谱波段，又能容忍光谱噪声。数据缩减将通过非参数技术，如样条拟合和惩罚判别分析（PDA）进行解释和分类目的。从最小描述长度（MDL）原则衍生的模型选择标准将作为一般模型选择工具和基于曲线数据的样条结选择和频带选择的标准进行研究。空间PDA将用于对来自机载光谱仪的高光谱图像进行分类，以考虑图像中像素的空间依赖性。正确认识森林物种在自然资源管理、环境保护、生物多样性和野生动物研究中具有重要意义。传统上可靠的树种识别方法主要依赖于野外调查或大尺度航空照片的判读。这些方法的使用经常受到成本和时间的限制，并且不适用于大面积。数字遥感已被用于大面积森林树种的识别，但遇到了两个问题：不同树种往往具有相似的光谱特征，部分原因是由于缺乏高光谱分辨率和大量的光谱波段；由于光学遥感所依据的光照条件不同，同一树种可能具有不同的光谱特性。

Bin YuDMS 9803063Hyperspectral data consist of intensity readings of hundreds of bands from ground or airborne spectrometers. The proposed research is to identify forest species by means of (1) hyperspectral data measured directly from above forest canopies in the field, and (2) hyperspectral data from airborne instruments based on the understanding from part (1). This research is motivated by the belief that the rich amount of spectral information contained in hyperspectral data should improve the level of discrimination of forest species and by the desire for a relatively simple method that can handle the large number of spectral bands in hyperspectral data while tolerant of spectral noise. Data reduction will be carried out via nonparametric techniques such as spline fitting and penalized discriminant analysis (PDA) for interpretation and classification purposes. Model selection criteria derived from the Minimum Description Length (MDL) principle will be investigated as both general model selection tools and criteria for spline knots selection and band selection for data reduction based on curve data. Spatial PDA will be developed to classify hyperspectral images from airborne spectrometers to take into account the spatial dependence of pixels in the image. Correct recognition of forest species is important in natural resource management, environmental protection, biodiversity and wildlife studies. Conventionally reliable methods for tree species recognition depend mainly on inventory in the field or on interpretation of large-scale aerial photographs. The use of these methods is frequently limited by cost and time and not applicable to large areas. Digital remote sensing has been used to identify forest species of large areas, but two problems are encountered: different tree species often have similar spectral characteristics partly due to the lack of high spectral resolution and lack of large number of spectral bands; and the same tree species may have distinct spectral properties due to illumination conditions on which optical remote sensing is based.