Fig. 7
An illustration of the Endmember Guided unmixing autoencoder (EGAE) method. The autoencoder is trained on an augmented dataset of 19 lakes, with epochs varying based on the objective function and hyperparameter optimization. The latent layer is split into two segments: one connected to trainable weights and the other to non-trainable, frozen weights initialized with reference spectra of pigments, i.e. chl-a, fx16, and pc17. Post-training, latent layer activations provide abundance estimates, and the first layer decoder weights yield endmember estimates with the non-trainable weights capturing background endmembers labeled as Endmember 1 to 6. The abundance estimates of the test set are derived by using only the encoder part and inputting the test dataset.
