A novel sequential feature-extraction approach for hyperspectral image classification using a recurrent neural network

Deep-learning (DL) algorithms have been introduced and applied in remote-sensing and other geospatially-related communities in recent years due to their significant capabilities and robustness compared with other traditional machine-learning models. Among them, recurrent neural networks (RNNs), which were initially designed to handle sequential data, have been investigated as classification models in recent years, particularly in the context of multi-temporal remote-sensing images. Regarding the single-image-based RNN classification models, extracting sequential features is the key problem, and two main strategies have been proposed to handle such an issue: pixel-feature-based sequential-feature representation, and similar pixels-based sequential-feature encoding, respectively. In this research, inspired by our previous work regarding similarity measurements-based sequential-feature construction, we propose a computationally-efficient sequential feature-extraction approach for the long short-term memory (LSTM)-based hyperspectral image (HSI) classification model. Within this model, an objected-based segmentation algorithm is applied first to generate a segmentation map. Then, similar segments are selected to guide similar-pixel searching from the whole-image scope to the segment-level scope. Experimental results derived from analysis of three benchmark HSI datasets demonstrate that accurate classifications are attained via the proposed methods, where distinctly lower computation-time cost is also achieved.