Improving Autoencoder Training Performance for Hyperspectral Unmixing with Network Reinitialisation

TL;DR

This paper investigates how autoencoder training for hyperspectral unmixing is affected by weight initialization and proposes reinitialization techniques to improve reconstruction accuracy and stability.

Contribution

It introduces network reinitialization methods based on neuron activation coefficients to enhance autoencoder training stability and performance in hyperspectral unmixing.

Findings

Reinitialization improves reconstruction accuracy.

Techniques reduce errors in abundances and endmembers.

Autoencoder training becomes more stable with proposed methods.

Abstract

Neural networks, in particular autoencoders, are one of the most promising solutions for unmixing hyperspectral data, i.e. reconstructing the spectra of observed substances (endmembers) and their relative mixing fractions (abundances), which is needed for effective hyperspectral analysis and classification. However, as we show in this paper, the training of autoencoders for unmixing is highly dependent on weights initialisation; some sets of weights lead to degenerate or low-performance solutions, introducing negative bias in the expected performance. In this work, we experimentally investigate autoencoders stability as well as network reinitialisation methods based on coefficients of neurons' dead activations. We demonstrate that the proposed techniques have a positive effect on autoencoder training in terms of reconstruction, abundances and endmembers errors.