Table 8.

Summary of review of HSI classification using deep learning—CNN.

Year	Method used	Dataset and COA	Research remarks and future scope
2015	Convolutional neural network and multilayer perceptron (CNN-MLP) [120]	Pavia city—99.91%, UP—99.62%, SV—99.53%, IP—98.88%	Far better than SVM, RBF mixed classifiers, the effective convergence rate can be useful for large datasets
			Detection of human behavior from hyperspectral video sequences
2016	3D-CNN [121]	IP—98.53%, UP—99.66%, KSC—97.07%	A landmark in terms of quality and overall performance
			Mapping performance to be accelerated by postclassification processing
2016	Spectral-spatial feature-based classification (SSFC) [122]	Pavia center—99.87%, UP—96.98%	Highly accurate than other methods
			Inclusion of optimal observation scale for improved outcome
2016	CNN-based simple linear iterative clustering (SLIC-CNN) [123]	KSC—100%, UP—99.64, IP—97.24%	Deals with a limited dataset use spectral and local-spatial probabilities as an enhanced estimate in the Bayesian inference
2017	Pixel-pair feature enhanced deep CNN (CNN-PPF) [124]	IP—94.34%, SV—94.8%, UP—96.48%	Overcomes the significant parameter and bulk-data problems of DL, PPFs make the system unique and reliable, and voting strategy makes the more enhanced evaluations in classification
2017	Multiscale 3D deep convolutional neural network (M3D-DCNN) [125]	IP—97.61%, UP—98.49%, SV—97.24%	Outperforms popular methods like RBF-SVM and combinations of CNNs
			Removing data limitations and improving the network architecture
2018	2D-CNN, 3D-CNN, recurrent 2D-CNN (R-2D-CNN), and recurrent 3-D-CNN (R-3D-CNN) [126]	IP-99.5%, UP—99.97%, Botswana—99.38%, PaviaC—96.79%, SV—99.8%, KSC—99.85%	R-3D-CNN outperforms all other CNNs mentioned and proves to be very potent in both fast convergence and feature extraction but suffers from the limited sample problem
			Applying prior knowledge and transfer learning
2019	3D lightweight convolutional neural network (CNN) (3D-LWNet) [127]	UP—99.4%, IP—98.87%, KSC—98.22%	Provides irrelevance to the sources of data
			Architecture is to be improvised by intelligent algorithms
2020	Hybrid spectral CNN (HybridSN) [128]	IP—99.75%, UP—99.98%, SV—100%	Removes the shortfalls of passing over the essential spectral bands and complex, the tedious structure of 2D-CNN and 3D-CNN exclusively and outruns all other contemporary CNN methods superiorly, like SSRN and M-3D-CNN
2020	Heterogeneous TL based on CNN with attention mechanism (HT-CNN-attention) [129]	SV—99%, UP—97.78%, KSC—99.56%, IP—96.99%	Efficient approach regardless of the sample selection strategies chosen
2020	Quantum genetic-optimized SR based CNN (QGASR-CNN) [27]	UP—91.6%, IP—94.1%	With enhanced accuracy, overfitting and “salt-and-pepper” noise are resolved
			Improvement of operational performance by the relation between feature mapping and selection of parameters
2020	Rotation-equivariant CNN2D (reCNN2D) [130]	IP—97.78%, UP—98.89, SV—98.18%	Provides robustness and optimal generalization and accuracy without any data augmentation
2020	Spectral-spatial dense connectivity-attention 3D-CNN (SSDANet) [131]	UP—99.97%, IP— 99.29%	Higher accuracy but high computational hazard
			Optimization by using other efficient algorithms