Spectral Feature Extraction Using Pooling

Hyperspectral imagery (HSI) has a huge number of narrow and contiguous bands. This involves high computation complexity in processing and analysing the image. Hence, dimensionality reduction is applied as an essential pre-processing step in hyperspectral data analysis. Pooling is a technique of reducing spatial dimension and is successfully applied in intermediate layers of convolutional neural networks for spatial feature extraction. There are various types of pooling strategies present, viz. max pool, mean pool, with their respective merits. In this chapter, the concept of pooling is applied in the spectral dimension of the hyperspectral data to reduce the dimensionality, and the results are compared with standard dimensionality reduction processes such as principal component analysis (PCA). Different pooling methods are applied and compared using two real hyperspectral datasets. The mean pooling technique is found to perform better among them. The results are compared in terms of overall accuracy and execution time. The experimental results show that the reduced signature patterns give an approximation of the actual spectral signatures of the original dataset, which is a very rare characteristic of feature extraction methods.