Table 6.

Summary of the review of HSI classification using deep learning—AE.

Year	Method used	Dataset and COA	Research remarks and future scope
2013	Autoencoders (AE) [110]	Error rate: KSC—4%, Pavia city—14.36%	This article opened a considerable doorway of research, including other deep models for better accuracy
2014	Stacked autoencoder and logistic regression (SAE-LR) [113]	KSC—98.76%, Pavia city—98.52%	Highly accurate in comparison to RBF-SVM and performs testing in optimized time limit than SVM or KNN but fails in training time efficiency
2016	Spatial updated deep AE with collaborative representation-based classifier (SDAE-CR) [114]	IP—99.22%, Pavia center—99.9%, Botswana—99.88%	Highly structured in extracting high specialty deep features and not the hand-crafted ones and accurate
			Improving the deep network architecture and selection of parameters
2019	Compact and discriminative stacked autoencoder (CDSAE) [115]	UP—97.59%, IP—95.81%, SV—96.07%	Efficient in dealing with feature space in low dimension, but the computation cost is high as per architecture size
2021	Stacked autoencoder with distance-based spatial-spectral vector [116]	SV—97.93%, UP—99.34%, surrey—94.31%	Augmentation of EMAP features with the geometrically allocated spatial-spectral feature vectors achieves excellent results. Better tuning of hyperparameter and more powerful computational tool required
			Improving the training model to become unified and classified in a more generalized and accurate way