Improving Autoencoder Training Performance for Hyperspectral Unmixing with Network Reinitialisation
This work addresses training stability issues in hyperspectral unmixing, which is crucial for accurate analysis and classification in remote sensing, but it is incremental as it builds on existing autoencoder methods.
The paper tackled the problem of autoencoder training instability for hyperspectral unmixing, showing that weight initialization significantly affects performance and proposing reinitialization methods based on dead neuron activations to reduce reconstruction, abundances, and endmembers errors.
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.