Spectral-Spatial Active Learning With Structure Density for Hyperspectral Classification

Qianming Li,Bohong Zheng,Yusheng Yang

doi:10.1109/access.2021.3074405

Abstract

In this paper, a spectral-spatial active learning (AL) method is proposed based on an up-to-date unlabeled samples sampling strategy concentrated on the structure density supported by breaking ties. The proposed sampling criterion in AL is used for hyperspectral image classification, which involves several steps: First, superpixel segmentation algorithm is conducted on the HSI to cluster pixels with similar spectral-spatial signature into the same superpixel block. Then, density peak clustering technique is performed on the each superpixel block to obtain structure density of the pixels. Meanwhile, probability-based classifier is employed to achieve the probability distributions of pixel. Next, breaking ties (BT) score of each pixel can be calculated by exploiting the probabilities. Additionally, a fusion mechanism is introduced to select the unlabeled samples with representativeness and informativeness advantages by employing the BT-assisted structure density (SD sampling criterion) of each pixel. Finally, the samples with manual labeled class labels are put into the training set to retrain the classifier. Experimental results manifest that the proposed SD-based sampling criterion in active learning can significantly improve the classification accuracy in few labor costs. Thus, it has certain feasibility in practical application.

Highlights

With the advance of spectroscopy sensor technology, hyperspectral image (HSI) with high spectral dimensionality and spatial resolution has been constantly becoming more available
EXPERIMENTAL DATASETS In order to verify the effectiveness of the proposed structure density (SD) sampling criterion, experiments are performed on four hyperspectral datasets1, i.e., Indian Pines scene, Salinas scene, University of Pavia scene, and Center of Pavia scene
1Datasets can be downloaded at: http://www.ehu.eus/ccwintco/index.php/ Hyperspectral_Remote_Sensing_Scenes resolution of 20 m per pixel (20 water absorption channels were removed before experiments)

Summary

Introduction

With the advance of spectroscopy sensor technology, hyperspectral image (HSI) with high spectral dimensionality and spatial resolution has been constantly becoming more available. Considering the abundance of spatial and spectral information, numerous classification algorithms exploiting remote sensing images have played a primary role in a variety of applications, such as precision agriculture [1], [2], environmental monitoring [3], land cover [4], and urban expansion [5]. A larger number of unlabeled samples with rich feature information did not play its role improving performance of classifier in supervised learning. In order to overcome the problem, many scholars and researchers in hyperspectral domain are willing to devote oneself to study advanced machine learning and classification methods, such as semi-supervised learning (SSL) [22], active learning (AL) [23], [24], semi-supervised active learning (SSAL) [25]–[28], and spectral-spatial classification [29]–[32]

Methods

Results

Conclusion