Abstract:

Aiming at the problem of limited classification accuracy due to insufficient mining of regular texture features in band selection of hyperspectral remote sensing images of crop planting areas， a band selection algorithm based on Mahalanobis distance and Gibbs-Markov Random Field （GMRF） spatial filtering was proposed. Firstly， for the regular texture features commonly found in crop planting areas， spatial filtering of hyperspectral images was performed by establishing a GMRF model， which retained and strengthened the texture features while reducing noise and redundant information， and enhanced the differences between ground object features. Then， a category separability metric was established on the basis of Mahalanobis distance combined with the ratio method， the contribution value of each band to the metric was calculated， and the bands were ranked according to the contribution values， thereby the specified number of top-ranked bands were selected as the output of the algorithm. The Indian Pines hyperspectral dataset， which contains a large number of crop planting areas， was used for band selection and maximum likelihood classification experiments， and the results show that compared with the optimal performance indexes of the three reference algorithms： genetic algorithm， successive projections algorithm， and density peak clustering algorithm， the proposed algorithm’s average correlation， overall classification accuracy and Kappa coefficient were improved by 3.37%， 2.90% and 6.70%， respectively. It can be seen that the proposed algorithm integrates crop spatial texture and spectral covariance features effectively， providing a feature selection scheme with clear physical interpretation for crop classification and growth monitoring in precision agriculture.

Key words: crop planting area, hyperspectral band selection, Mahalanobis distance, Gibbs-Markov Random Field (GMRF), spatial filtering, category separability

摘要：

针对作物种植区域的高光谱遥感图像波段选择中规则纹理特征挖掘不足导致分类精度受限的问题，提出一种基于马氏距离结合吉布斯-马尔可夫随机场（GMRF）空间滤波的波段选择算法。首先，针对作物种植区域普遍存在的规则纹理特征，通过建立GMRF模型对高光谱图像进行空间滤波，从而在降低噪声和冗余信息的同时保留并强化纹理特征，并增强地物之间的特征差异；其次，基于马氏距离和比值法建立类别可分性度量指标，计算每个波段对该指标的贡献值，并依据贡献值大小进行波段排序，从而选择排序靠前的指定数量波段作为算法输出。采用包含大量作物种植区域的Indian Pines高光谱数据集进行波段选择和最大似然分类实验，结果表明，相较于遗传算法、连续投影法和峰值聚类法这3种参考算法中的最优性能指标，该算法的平均相关性、总体分类精度和Kappa系数分别改善了3.37%、2.90%和6.70%。可见，该算法有效融合了作物空间纹理与光谱协方差特征，为精准农业中的作物分类与长势监测提供了具有明确物理解释性的特征选择方案。

关键词: 作物种植区域, 高光谱波段选择, 马氏距离, 吉布斯-马尔可夫随机场, 空间滤波, 类别可分性