The use of hyperspectral satellite data (e.g., PRISMA and EnMAP) to feed Linear Spectral Mixture Models has proven effective in improving LC classification but showed a challenge in the classification of mixed pixels, characteristic of transitional environment. In particular, the selection of appropriate Fractional Abundance Maps (FAMs) threshold continues to represent a critical step in the spectral unmixing process. To address the challenge of discretizing biophysical gradients, the data processing workflow leverages Active Learning techniques and field-based spectral libraries to fill this gap.