基于植被特征库的高光谱植被精细分类

摘要

There are two major problems of sophisticated vegetation classification (SVC) using hyperspectral image .Classification results using only spectral information can hardly meet the application requirements with the needed vegetation type becoming more sophisticated .And applications of classification image are also limit‐ed due to salt and pepper noise .Therefore the SVC strategy based on construction and optimization of vegeta‐tion feature band set (FBS) is proposed .Besides spectral and texture features of original image ,30 spectral in‐dices which are sensitive to biological parameters of vegetation are added into FBS in order to improve the sepa‐rability between different kinds of vegetation .And to achieve the same goal a spectral‐dimension optimization algorithm of FBS based on class‐pair separability (CPS) is also proposed .A spatial‐dimension optimization al‐gorithm of FBS based on neighborhood pixels’ spectral angle distance (NPSAD) is proposed so that detailed information can be kept during the image smoothing process .The results of SVC experiments based on air‐borne hyperspectral image show that the proposed method can significantly improve the accuracy of SVC so that some widespread application prospects like identification of crop species ,monitoring of invasive species and precision agriculture are expectable .%目前，高光谱数据精细分类面临两方面问题：一方面，传统单纯利用光谱信息的分类往往难以满足各应用行业对精度的需求，另一方面，基于像元的分类结果受制于椒盐噪声，影响其有效应用。为此，提出了一种基于植被特征库构建与优化的高光谱植被精细分类策略。首先，从高光谱影像中的原始光谱特征出发，结合灰度共生矩阵和局域指示空间分析两类纹理特征，并有针对性地加入了对植被叶绿素、胡萝卜素、花青素和氮素叶面积指数等理化参量敏感的光谱指数特征，构建了完备的植被特征库，以提高植被类别间的可分性；进而对植被特征库进行光谱维优化，提出了基于类对可分性的光谱维优化算法，选择对各类对具有最高识别能力的特征波段，通过迭代使各类别间均达到较高的区分度，并利用最优索引因子法进一步降低数据冗余，以提高分类效率；在进行植被特征库空间维优化时，主要基于地物分布通常具有一定的空间连续性这一理论，提出了基于邻域光谱角距离的植被特征库空间维优化算法，以去除分类结果中的椒盐噪声，提高分类精度和分类图像平滑度。基于航空高光谱数据的植被精细分类验证表明，该方法可以显著提高分类精度，在作物品种识别、精准农业等方面将具有广泛的应用前景。