Table 12.

Summary of review of HSI classification using transfer learning.

Year	Method used	Dataset and COA	Research remarks and future scope
2018	Deep mapping-based heterogeneous transfer learning model (DLTM) [150]	Washington DC Mall—96.25%	Capable of binary classification
			Improvisation to multiclass classification
2018	AL with stacked sparse autoencoder (AL-SSAE) [151]	UP—99.48%, center of Pavia—99.8%, SV— 99.45%	Domains, both source, and target possess finely tuned hyperparameters
			Architectural parameters need to be modified further to enhance the classification accuracy
2020	Heterogeneous TL based on CNN with attention mechanism (HT-CNN-attention) [152]	SV—99%, UP—97.78%, KSC—99.56%, IP—96.99%	Efficient approach regardless of the sample selection strategies chosen
2020	ELM-based ensemble transfer learning (TL-ELM) [26]	UP—98.12%, Pavia center—96.25%	Efficient accuracy and transferability with high training speed
			Inclusion of SuperPCA and knowledge transfer
2020	Lightweight shuffled group convolutional neural network (abbreviated as SG-CNN) [153]	Botswana—99.67%, HU—99.4%, Washington DC—97.06%	Fine-tuned model as compared to CNN architectures, low computational cost for training
			Inclusion of more grouped convolutional architectures
2021	Super-pixel pooling convolutional neural network with transfer learning (SP-CNN) [154]	SV—95.99%, UP—93.18%, IP—94.45%	More excellent parameter optimization with more accuracy using a limited number of samples and in a very short period for both training and testing
			Optimal super-pixel segmentation and merging with different CNN architectures