Multiattribute Sample Learning for Hyperspectral Image Classification Using Hierarchical Peak Attribute Propagation

Although hyperspectral image (HSI) classification methods have gained popularity in the remote sensing community, their performance tends to be limited by the quantity and quality of training samples. Actually, the production (i.e., capture and annotation) of training samples requires abundant labor costs and time consumption. Multiattribute sample learning can effectively solve the cost problem for the acquisition of training samples; for this, this article proposes a novel hierarchical peak attribute propagation (HPAP) method for the annotation of multiattribute samples, which is composed of the following key technologies. First, superpixel-oriented guided filtering (SOGF) is designed to improve the distinguishability of spectral features between pixels. Then, peak attribute-driven manual labeling is performed on each multiattribute region to provide prior knowledge for label propagation. Next, peak attribute-based label propagation is applied to each multiattribute region to achieve the single-label information of pixels. Finally, multiattribute samples with a single class label are augmented to the training process of HSI supervised model to enhance their classification performance. Experimental results on several public HSI datasets indicated that the proposed HPAP method can obtain higher annotation accuracy for multiattribute samples and promote the classification accuracy of HSI classification supervised models.