A Metareinforcement-Learning-Based Hyperspectral Image Classification With a Small Sample Set

Abstract

The fine spectral information contained in hyperspectral images (HSI) is combined with rich spatial features to provide feature qualities that serve as distinguishing variables for efficient classification performance. The task's objective is to correctly identify and categorize several object categories in the HSI, such as the ground, flora, water, and buildings, based on their spectral characteristics beneficial for a variety of applications, including mapping minerals, analyzing vegetation, and mapping urban land-use. The difficulty of learning new taskspecific knowledge from a limited data sample that encourages less training has not been overcome by deep learning models. The capacity of current models to generalize to new tasks on small data sets is still lacking. By learning features that are transferable to facilitate adaptation to novel tasks on small samples, meta-reinforcement learning shows promise in overcoming such difficulties. We proposed a meta-reinforcement learning (Meta-RL) model that decouples task inference to improve metatraining, and accelerate meta-learning with small HSI labeled samples for efficient classification. The model employs a Capsule network for effective cooperation between spectra-spatial bands. To minimize the temporal difference error, the Apex-X Deep Q network parameter update is used to meta-train our model. The proposed model obtains an overall accuracy between 95.85% and 96.78% with computational time between 3207.9s and 7487.9s for training and validation as well as between 21.57s and 32.98s for testing. The experimental results prove the competitiveness of the proposed model to existing traditional deep learning, meta-learning, and reinforcement learning methods in both classification accuracy and computational cost.