Authors: Andong Ma*, Texas A&M University, Anthony M. Filippi, Texas A&M University, Burak Güneralp, Texas A&M University
Topics: Remote Sensing, Geographic Information Science and Systems
Keywords: Hyperspectral image (HSI) classification, long short-term memory (LSTM), recurrent neural network (RNN), similarity measurements
Session Type: Virtual Paper
Start / End Time: 4:40 PM / 5:55 PM
Room: Virtual 43
Presentation File: No File Uploaded
Deep learning (DL), a sub-field of artificial intelligence where “deeper” and highly-abstracted features are extracted and encoded, relative to conventional methods, has attracted increasing attention in the fields of computer vision and image processing. Among well-known DL algorithms, recurrent neural networks (RNNs), which have distinguished adaptability and capability for handling sequence data, have drawn increasing attention in the field of remote sensing, including hyperspectral remote-sensing image classification. A typical application domain for RNNs is multi-temporal remote-sensing image classification, due to the intrinsic sequential features in such images collected at different times. However, for the single-image classification task, extracting such sequential features from an individual image is still a critical issue. This study, inspired by our previous work regarding similarity-based sequential-feature construction, proposes an improved sequential feature-extraction framework with a lower computational time cost. Within this framework, an objected-based segmentation algorithm is applied first to generate the segmentation maps. Then, both local and non-local segments are considered for pixel-wise similarity calculations. Unlike our previous work, where similar-pixel searching is implemented on the whole image, such a search range is reduced to the segments set in the present research. Experimental results derived from three standard hyperspectral images (with corresponding reference datasets) demonstrate that more accurate classification performances are obtained via the proposed methods, where distinctly lower computational time cost is achieved as well.