Table 2.

Comparison of six band selection methods.

	Principle	Advantages	Disadvantages	Differences and Similarities
Ranking-based	Use a suitable function to quantify the amount of information in each band, and then select the top subset of bands according to their importance	Low computational complexity and fast execution of calculations for larger hyperspectral datasets	Correlation between bands is often not considered	Search-based, sparsity-based, and embedding-learning band selection methods are all optimization problems with objective functions; ranking-based and clustering-based band selection methods are all based on the importance of bands. And all band selection methods are designed to select the combination of bands with high information content, low correlation between bands, and best class separability.
Search-based	The optimization problem of the criterion function is a multi-objective optimization to find the optimal frequency band	Only individual bands are considered, ignoring the entire subset of bands optimized	Computationally intensive and difficult to apply in practice
Clustering-based	The representative subset of frequency bands in the cluster of the component group	Entire subset of bands can be optimized; less affected by noise; simple algorithm	Poor robustness, easy to fall into local optimal solutions
Sparsity-based	Obtaining representative bands by dealing with sparsely constrained optimization problems	Can reduce the complexity of hyperspectral data processing; reduce storage space; improve model interpretability	Difficulty in automating model applications; uncertainty in model processing performance
Embedded learning-based	Optimize the objective function of a specific model and select the appropriate spectral band	Avoids repetitive training of the learner for each subset of bands	Performance-dependent parameter tuning and difficult objective function construction
Hybrid scheme-based	A synthesis of several band selection algorithms	Can find the best combination of frequency bands to get the least number of useful bands	Algorithm complexity