Table 5.

Summary of review of HSI classification using active learning.

Year	Method used	Dataset and COA	Research remarks and future scope
2008	AL with expectation-maximization-binary hierarchical classifier (BHC-EM-AL) and maximum-likelihood (ML-EM-AL) [90]	Range: KSC-90-96%, Botswana—94-98%	Better learning levels than the random choice of data points and an entropy-based AL
			Measurement of the efficacy of the active learning-based knowledge transfer approach while systematically increasing the spatial/temporal segregation of the data sources
2010	Semi-supervised-segmentation with AL and multinomial logistic regression (MLR-AL) [91]	IP—79.90%, SV—97.47%	Innovative mechanisms for selecting unlabeled training samples automatically, AL to enhance segmentation results
			Testing the segmentation in various scenarios influenced by limited a priori accessibility of training images
2013	Maximizer of the posterior marginal by loopy belief propagation with AL (MPM-LBP-AL) [92]	IP—94.76%, UP—85.78%	Improved accuracy than previous AL applications
			Use parallel-computer-architectures such as commodity—clusters or GPUs to build computationally proficient implementation
2015	Hybrid AL-MRF, that is, uncertainly sampling breaking ties (MRF-AL-BT), passive selection approach random sampling (MRF-AL-RS), and the combination (MRF-AL-BT + RS) [93]	IP—94.76%, UP—85.78% (MRF-AL-RS provides the highest accuracies)	Outperforms conventional AL and SVM AL methods due to MRF regularization and pixelwise output
			Merge the model with other effective AL methods and test them with a limited number of training samples
2015	Integration of AL and Gaussian process classifier (GP-AL) [94]	IP—89.49%, Pavia center—98.22%	Empirical autonomation of AL achieves reasonable accuracy
			Adding diversity criterion to the heuristics and contextual information with the model and reducing computation time
2016	AL with hierarchical segmentation (HSeg) tree: adding features and adding samples (Adseg_AddFeat + AddSamp) [95]	IP—82.77%, UP—92.23%	Outruns several baseline methods-selecting appropriate training data from already existing labeled datasets and potentially decreasing manual laboratory labeling
			Reduce the computational time that limits its applicability on large-scale datasets
2016	Multiview 3D redundant discrete wavelet transform-based AL (3D-RDWT-MV-AL) [96]	HU—99%, KSC—99.8%, UP—95%, IP—90%	The precious method as a combination of an initial process with AL, improved classification
2017	Discovering representativeness and discriminativeness by semi-supervised active learning (DRDbSSAL) [97]	Botswana—97.03%, KSC—93.47%, UP—93.03%, IP—88.03%	Novel approach with efficient accuracy
2017	Multicriteria AL [98]	KSC—99.71%, UP—99.66%, IP—99.44%	Surpasses other existing AL methods regarding stability, accuracy, robustness, and computational hazard
			A multi-objective optimization strategy and the usage of advanced attribute-based profile features
2018	Feature-driven AL associated with morphological profiles and Gabor filter [99]	IP—99.5%, UP—99.84%, KSC—99.53% (Gabor-BT)	A discriminative feature space is designed to gather helpful information into restricted samples
2018	Multiview intensity-based AL (MVAL)-multiview intensity-based query-representative strategy (MVIQ-R) [100]	UP—98%, Botswana—99.5%, KSC—99.9%, IP—95%	Focus on pixel intensity obtains unique feature and hence better performance
			Selection of combination of optimal attribute features
2019	Super-pixel with density peak augmentation (DPA)-based semi-supervised AL (SDP-SSAL) [101]	IP—90.08%, UP—85.61%	Novel approach proposed based on super-pixels density metric
			Development of a pixelwise solution to produce super-pixel-based neighborhoods
2020	Adaptive multiview ensemble spectral classifier and hierarchical segmentation (Ad-MVEnC_Spec + Hseg) [102]	KSC—97.63%, IP—87.1%, HU—93.3%	Enhancement in the view sufficiency, and promotion of the disagreement level by the dynamic view, provides lower computational complexity due to parallel computing
2020	Spectral-spatial feature fusion using spatial coordinates-based AL (SSFFSC-AL) [103]	IP—100%, UP—98.43%	High running speed can successfully address the “salt and pepper” phenomenon but drops a few if similar class samples are distributed in different regions differently
			The sampling weight parameter conversion to an adaptive parameter is adjusted adaptively as the training samples are modified