Table 9.

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

Year	Method used	Dataset and COA	Research remarks and future scope
2017	Gated recurrent unit-based RNN with parametric rectified tanh as activation function (RNN-GRU-pretanh) [132]	UP—88.85%, HU—89.85%, IP—88.63%	An enhanced model that utilizes the intrinsic feature provided by HS pixels with better accuracy than SVM
			The study is limited to only spectral features
			Incorporation of deep end-to-end convolutional RNN with both spatial-spectral features
2019	Spectral-spatial cascaded recurrent neural network (SSCasRNN) [135]	IP—91.79%, UP—90.30%	Outruns pure RNN and CNN models due to the perfect placement of convolutional and recurrent layers to explore joint information
2020	Geometry-aware deep RNN (Geo-DRNN) [136]	UP—98.05%, IP—97.77%	Due to encoding the complex geometrical structures, the data lack space
			Minimization of memory-occupation
2021	2D and 3D spatial attention-driven recurrent feedback convolutional neural network (SARFNN) [28]	IP—99.15%, HU—86.05%	Integrating attention and feedback mechanism with recurrent nets in two layers, 2D and 3D, enables efficient accuracy