HYPERSPECTRAL IMAGE CLASSIFICATION VIA OBJECT-ORIENTED SEGMENTATION-BASED SEQUENTIAL FEATURE EXTRACTION AND RECURRENT NEURAL NETWORK

Recurrent neural networks (RNNs) have been investigated and utilized as classification model in the hyperspectral remote-sensing community due to its great capability of encoding sequential features, especially for multi-temporal images. For non-temporal, individual remote-sensing images, RNNs are still a dominant and powerful classification tool that benefits from sequential feature extraction from a single image. In this article, we propose a computationally-efficient sequential feature extraction method for the long short-term memory (LSTM)-based hyperspectral image classification model. Within the proposed method, object-oriented segmentation was employed first to guide similar-pixel searching in the whole-image scope to a local segment scope. Experimental results on two benchmark hyperspectral datasets indicate that our proposed methods achieve higher classification accuracy with lower computational cost.