Spectral-Spatial Hyperspectral Image Semisupervised Classification by Fusing Maximum Noise Fraction and Adaptive Random Multigraphs

Table 1

A comparison of different methods for hyperspectral image classification.

Group	Subgroup	Method	Advantages	Difficulties
Segmentation-based methods	Object-based classification	Pixon-based classifier	The noise pixels in the classification map were removed and the suitable land cover smoothing map was obtained	The spectral characteristics of abnormal pixels are often similar to the background, resulting in the loss of information when the abnormal pixels are removed
Feature fusion	Features stacking	PCA-SPCA-2D-SSA	Combined with appropriate spatial features, the classification efficiency of the algorithm will be higher	Complex structure
	Joint spectral-spatial FE	R-VCANet	The use of high correlation among SSI	Complex structure
	Meta-learning-based classifiers	RN-FSC	Less sensitive to the number of training samples	Limited generalization ability for large-scale hyperspectral datasets
	Deep learning-based classifiers	iCapsNet	Well-initialized shallow layers	Complex structure
		RPNet	FE and classification are carried out under a unified framework	Insufficient training samples will lead to overfitting
Decision fusion	Decision fusion	MBFSDA	Combination of supplemental information and several advanced classifiers	Selection of suitable feature extractor
		Our method	(i) Self-adaptability of parameters; (ii) Leveraging tiny labeled samples	Complex structure