Deep learning for hyperspectral image classification: A survey

Abstract

Hyperspectral image (HSI) classification is a significant topic of discussion in real-world applications. The prevalence of these applications stems from the precise spectral information offered by each pixelś data in hyperspectral imaging (HS). Classical machine learning (ML) methods face challenges in precise object classification with HSI data complexity. The intrinsic non-linear relationship between spectral information and materials complicates the task. Deep learning (DL) has proven to be a robust feature extractor in computer vision, effectively addressing nonlinear challenges. This validation drives its integration into HSI classification, which proves to be highly effective. This review compares DL approaches to HSI classification, highlighting its superiority over classical ML algorithms. Subsequently, a framework is constructed to analyze current advances in DL-based HSI classification, categorizing studies based on a network using only spectral features, spatial features, or both spectral–spatial features. Moreover, we have explained a few recent advanced DL models. Additionally, the study acknowledges that DL demands a substantial number of labeled training instances. However, obtaining such a large dataset for the HSI classification framework proves to be time and cost-intensive. So, we also explain the DL methodologies, which work well with the limited training data availability. Consequently, the survey introduces techniques aimed at enhancing the generalization performance of DL procedures, offering guidance for the future.