Table 3.

Summary of review of HSI classification using MRF.

Year	Method used	Dataset and COA	Research remarks and future scope
2011	Adaptive-MRF (a-MRF) [59]	IP—92.55%	Handles homogeneous problem of “salt and pepper” areas and the possibility of overcorrection impact on class boundaries
2014	Hidden MRF and SVM (HMRF-SVM) [60]	IP—90.50%, SV—97.24%	Outperforms SVM and improves overall accuracy outcomes by nearly 8% and 3.2%, respectively
2014	Probabilistic SR with MRF-based multiple linear logistic (PSR-MLL) [61]	IP—97.8%, UP—99.1%, Pavia center—99.4%	Exceeds other modern contemporary methods in terms of accuracy
2014	MRF with Gaussian mixture model (GMM-MRF) [62]	UP(LFDA-GMM-MRF)-90.88% UP(LPNMF-GMM-MRF)—94.96%	Advantageous for a vast range of operating conditions and spatial-spectral information to preserve multimodal statistics
			GMM classificatory distributions are to be considered in the future
2011	MRF with sparse multinomial logistic regression classifier—spatially adaptive total variation regularization (MRF-SMLR-SpATV) [63]	UP—90.01%, IP—97.85%, Pavia center—99.23%	Efficient time complexity of the model
			Improvisation of the model by implementing GPU and learning dictionaries are the future agendas
2016	Multitask joint sparse representation (MJSR) and a stepwise Markov random filed framework (MSMRF) [64]	IP—92.11%, UP—92.52%	The gradual optimization explores the spatial correlation, which significantly improves the effectivity and accuracy of the classification
2016	MRF with hierarchical statistical region merging (HSRM) [65]	SVMMRF-HSRM: IP—93.10%, SV—99.15%, UP— 86.52%; MLRsubMRF-HSRM-IP—82.60%, SV—88.16%, UP—95.52%	Better solution to the technique of majority voting that suffers from the problem of scale choice
			Considering the spatial features in the spatial prior model of objects of the different groups in the future
2018	Integration of optimum dictionary learning with extended hidden Markov random field (ODL-EMHRF) [66]	ODL-EMHRF-ML-IP—98.56%, UP—99.63%; ODL-EMHRF-EM-IP—98.47%, UP—99.58%	The method has been proven to be better than SVM-associated EMRF
2018	Label-dependent spectral mixture model (LSMM) fused with MRF (LSMM-MRF) [67]	The Konka image—94.19%, the shipping scene—66.45%	Efficient unsupervised classification strategy that considers spectral information in mixed pixels and the impact of spatial correlation
			Enhanced theoretical derivations of EM steps
2019	Adaptive interclass-pair penalty and spectral similarity information (aICP2-SSI) along with MRF and SVM [68]	UP—98.10%, SV—96.40%, IP— 96.14%	Outperforms other MRF-based methods
			More efficient edge-preserving strategies, more spectral similitude, and class separable calculation methods as future research
2019	Cascaded version of MRF (CMRF) [69]	IP—98.56%, Botswana—99.32%, KSC—99.24%	Backpropagation tunes the model parameters and least computation expenses
2020	Fusion of transfer learning and MRF (TL-MRF) [70]	IP—93.89%, UP—91.79%	TL is taken to be very effective for HSI classification
			Future research for reducing the number of calculations involved in the existing
2020	MRF with capsule net (caps-MRF) [71]	IP—98.52%, SV—99.74%, Pavia center—99.84%	Ensures that relevant information is preserved, and the spatial constraint of the MRF helps achieve more precise model convergence
			The combination of CapsNet with several postclassification techniques