Hyperspectral Image Dimension Reduction and Target Detection Based on Weighted Mean Filter and Manifold Learning

Abstract

Abstract Hyperspectral images have many bands, resulting in a high data volume, which complicates subsequent data processing. Using manifold learning methods to reduce the dimension of data is conducive for subsequent use. However, traditional manifold learning methods are easily disturbed by spectral uncertainty, and are primarily used for hyperspectral image classification. This paper proposes an improved manifold reconstruction preserving embedding algorithm based on weighted mean filter (WMF-IMRPE), that is not easily affected by spectral uncertainty and has excellent target detection performance. The original hyperspectral image is processed by the weighted mean filter to eliminate the influence of noise and reduce spectral differences between the homogeneous ground objects. The spectral angular distance then replaces the Euclidean distance in the original MRPE algorithm to select neighbourhood pixels, reducing spectral uncertainty interference. The experimental results show that the low dimensional features extracted by the WMF-IMRPE algorithm have better distinguishability, and the algorithm further improves the target detection accuracy. The WMF-IMRPE algorithm’s hyperspectral image target detection performance is superior to other similar algorithms.