Exploring traditional, deep learning and hybrid methods for hyperspectral image classification: A review

Hyperspectral image (HSI) classification has appeared as an important area in remote sensing due to the rich spectral and spatial information that enables accurate object classification. Traditional methods such as sparse representation , random forest , support vector machines , and Markov random fields have long been used for HSI classification, offering remarkable efficiency and precision. Regardless, the ingrained complexity and nonlinearity of HSI data pose challenges to these traditional methods. Deep learning (DL) methods, including convolutional neural networks , recurrent neural networks , autoencoders , generative adversarial networks , and transformers, have shown super performance by leveraging their ability to capture both the spatial and spectral features separately or together. In recent years, hybrid methods that combine traditional methods and DL algorithms or integrate two types of DL algorithms have demonstrated superior performance by utilizing the strengths of different methods. This paper presents a comprehensive review of traditional approaches, DL approaches , and hybrid approaches in HSI classification and an overview of hyperspectral datasets. In addition, we compare these methods via evaluation metrics , highlighting the strengths and limitations of each type. Our work provides helpful guidelines for future research directions and work in this area.