Journal of Remote Sensing & GIS

Abstract

Machine Learning-based Estimation of the Number of Endmembers for Unmixing Hyperspectral Image

Bipasha Chakrabarti*

Spectral unmixing involves understanding the ground scene by inferring the endmember reflectance pattern, and computing their respective fractional abundance. Unmixing methods can be categorized as blind or semi-blind approach, based on the availability of spectral library data. Many existing methods for unmixing and endmember determination assume prior knowledge of the number of endmembers in the image scene. However, in reality, the number of endmembers is mostly unknown, besides, considering a huge sized spectral library as the endmember set leads to specific predicaments. Therefore, proper estimation of the number of consistent materials or endmembers is a vital task. The unmixing methods tend to consider mixed pixels as endmembers in case of overestimation of endmember number. On the other hand, some actual endmembers are unidentified due to underestimation. The eigenvalue of the covariance matrix of the Hyper Spectral Image (HIS) data points out to the number of endmembers, which is a specific numerical rank identification task. However, since the eigenvalue pattern itself gets modified due to the presence of noise, small sample size, inferring the number is a challenging task. Instead of the traditional approaches, this work formulated the task in terms of supervised learning, where the machine learning method learns the eigenvalue pattern for known number of endmembers. For this purpose, we created a dataset by cropping added noise to the existing hyperspectral datasets with well-known unmixing ground-truth. Next, we trained the machine learning models with the eigenvalue pattern, and the endmember number as the class. As per extensive experiments on several real hyperspectral datasets, the proposed network outperforms the other state-of-the-art methods and machine learning approaches.

Published Date: 2023-12-25; Received Date: 2023-11-22