226 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE DECEMBER 2021 Estimate Abundances and EMs for Each Pixel Color Code for the Computational Cost of the Unmixing Methods: Yes Are Spectral Libraries Available a Priori? Yes No Is Expert Knowledge Available? Library-Based Spectral Transformation Yes Yes Apply Spectral Transformation to the Image and Library? User-Defined Spectral Transformation No Is the Library Very Large? Yes Less Supervision Less Supervision or Less Computational Cost? Fuzzy Unmixing MESMA and Variants Less Cost Yes Bayesian Models No EM Libraries Extraction Extract Libraries From the Observed Image? Less Supervision No No Low Medium High Less Supervision or Less Computational Cost? Less Cost Prune Signatures to Make the Library Small? No Parametric EM Models EM-ModelFree Methods Local Unmixing Sparse Unmixing Machine Learning Methods FIGURE 2. The decision tree for hyperspectral unmixing, considering spectral variability. The blue boxes denote families of unmixing algorithms, while the yellow boxes represent additional techniques related to the extraction and processing of spectral libraries. MESMA: multiple-EM spectral mixture analysis.