Joint Posterior Probability Active Learning for Hyperspectral Image Classification

Active learning (AL) is an approach that can reduce the dependence on the labeled set significantly. However, most current active-learning methods are only concerned with the first two columns of the posterior probability matrix during the sampling phase. When the difference between the first and second-largest posterior probabilities of several samples is proximate, these approaches fail to distinguish them further. To improve these deficiencies, we propose an active learning algorithm, joint posterior probabilistic active learning combined with conditional random field (JPPAL_CRF). In the active-learning sampling phase, a new sampling decision function is built by jointing all the information in the posterior probability matrix. By doing so, the variability between different samples is refined, which makes the selected samples more meaningful for classification. Then, a conditional random field (CRF) approach is applied to mine the regional spatial information of the hyperspectral image and optimize the classification results. Experiments on two common hyperspectral datasets validate the effectiveness of JPPAL_CRF.

Automated summary

Active learning is a method that reduces the dependence on labeled samples. However, current active learning methods often fail to distinguish between samples with similar posterior probabilities. To address this issue, a novel algorithm called Joint Posterior Probabilistic Active Learning combined with Conditional Random Field (JPPAL_CRF) is proposed. This algorithm improves the variability between samples by jointly considering all information in the posterior probability matrix, and utilizes a conditional random field approach to optimize classification results by incorporating regional spatial information of hyperspectral images. Experimental results on two common hyperspectral datasets demonstrate the effectiveness of JPPAL_CRF.