利用区域增长技术的自适应高光谱图像分类

针对面向对象的高光谱分类方法中分割参数设置问题,提出了一种基于区域增长技术的自适应高光谱分类算法。首先提出了带约束的区域增长方法,利用已知训练样本的空间信息,提供有效约束,从而降低区域增长过程中区域标记的错误传播率,以提高分类性能;其次,提出了自适应阈值计算方法,通过分析已知训练样本光谱的分布规律,自适应地计算出合理的区域划分阈值,从而代替经验阈值,提高算法的鲁棒性;最后,采用K近邻算法(KNN),对划分后各区域中心进行分类。实验结果表明:对于不同图像,提出的算法计算出的自适应阈值均与其经验值相符合,且其分类效果优于其他算法,来自AVIRIS传感器的高光谱数据Indian Pines在10%的已知训练样本下总体分类精度达92.94%、kappa系数达0.919 5,来自ROSIS传感器的高光谱数据Pavia University在5%的已知训练样本下总体分类精度达95.78%、kappa系数达0.944 0。该算法不仅增强了算法的鲁棒性,同时有效提高了分类性能,在高光谱应用中具有较强的实用性。

Adaptive hyperspectral image classification using region-growing techniques

Aiming at the problem of segmentation parameters setting in object-oriented hyperspectral classification method, an adaptive hyperspectral classification algorithm based on region-growing techniques was proposed in this paper. Firstly, a constrained region-growing method was proposed, which used the spatial information of the training samples to provide effective constraints, thus reducing the error propagation rate of the region markers in the region-growing process, and improving classification performance. Secondly, an adaptive threshold calculation method was proposed. By analyzing the distribution law of the spectrum of the training samples, the reasonable threshold for region division was calculated adaptively to replace the empirical threshold, so that the robustness of the algorithm was improved. Finally, the K-nearest neighbor algorithm (KNN) was used to classify the centers of each region after division. Experimental results show that:For different images, the adaptive thresholds calculated by the method are consistent with the empirical values, and the classification effect of the proposed algorithm is better than other algorithms. For hyperspectral data Indian Pines from AVIRIS sensor, the overall classification accuracy and kappa are 92.94% and 0.919 5 respectively with 10% training samples, and for hyperspectral data Pavia University from ROSIS sensor, the overall classification accuracy and kappa are 95.78% and 0.944 0 respectively with 5% training samples. The proposed algorithm not only enhances the robustness of the algorithm, but also improves the classification performance effectively, and has strong practicability in hyperspectral applications.