Active learning-based hyperspectral image classification: a reinforcement learning approach

In the last few years, deep neural networks have been successful in classifying hyperspectral images (HSIs). However, training deep neural networks needs a large number of labeled datasets. In HSIs, acquiring a large amount of labeled data is costly and time-consuming. Active learning (AL) is a technique for selecting a small subset of data for annotation so that the classifier can learn from the data with high accuracy. Most of the AL methods are designed based on some statistical approach. The efficacy of the statistical methods is limited, and their performance varies depending on the scenario. So, a reinforced pool-based deep active learning (RPDAL) approach is proposed to overcome limitations of statistical selection approaches. The reinforcement learning (RL)-based agent is designed and trained to select informative samples for annotation. The learned RL-based agent can transfer and choose samples for annotation on any other HSI dataset after being trained on one. Indian Pines (IP), Pavia University (PV), and Salinas Valley (SL) are three publicly available datasets used in the experiment. The proposed approach achieves 92.78%, 97.85%, and 97.94% accuracy using 400 labeled samples with IP, PV, and SL datasets, respectively. The labeled samples selected using the proposed approach achieve better classification performance than other AL techniques.