Deep Reinforcement Learning for Band Selection in Hyperspectral Image Classification

Band selection refers to the process of choosing most relevant bands in a hyperspectral image. By selecting limited number optimal bands, we aim at speeding up model training, improving accuracy, or both. It reduces redundancy among spectral while trying preserve original information now, many efforts have been made develop unsupervised band approaches, which majorities are heuristic algorithms devised by trial and error. In this article, interested training an intelligent agent that, given image, is capable automatically learning policy select subset without any hand-engineered reasoning. To end, frame problem as Markov decision process, propose effective method parameterize it, finally solve deep reinforcement learning. Once trained, it learns band-selection that guides sequentially fully exploiting image previously picked bands. Furthermore, two different reward schemes for environment simulation compare them experiments. This, best our knowledge, first study explores analysis, thus opening new door future research showcasing great potential remote sensing applications. Extensive experiments carried out on four data sets, experimental results demonstrate effectiveness proposed method. The code publicly available.