Matrix Cofactorization for Joint Representation Learning and Supervised Classification -- Application to Hyperspectral Image Analysis

This code implements the method described in

Lagrange, A., Fauvel, M., May, S., Bioucas-Dias, J., & Dobigeon, N. (2020). Matrix cofactorization for joint representation learning and supervised classification–Application to hyperspectral image analysis. Neurocomputing, 385, 132-147.

Disclaimer

Feel free to contact me if you encounter any problem with the code.

Usage

An exemple of how to use the code is provided in the file runExpl.py.

Class initialization inputs

JointClassifUnmixModel(lambdas=[1.,1.,1.,0.1,0.],lambda_q=0.1,sigma=0.01,epsilon=0.01,tol=1e-4)

• lambdas: hyperparamaters weighting the terms of the objective function: 0)Y-MA 1)A-BZ 2)CD-QZD 3)||A||_1 4)||Z||_TV (cf paper)
• lambda_q: hyperparameter weighting the penalization of matrix Q, used only when classif_loss is 'cross-entropy'
• sigma: parameter in weight of adaptative TV (cf paper)
• epsilon: parameter to make the TV-norm gradient-Lipschitz (cf paper)
• tol: tolerance used to check convergence

I/O of optimization method

NB: be careful with labels: label 0 in input is equivalent to a unknown label

JointClassifUnmixModel.optimize(Y,R,K,C,labels,maxIter,classif_loss='cross-entropy',M=None,maxInitIter=1000,adaptTV=False,printEnable=False,initClassif=False,initFilePath=None,srand=0)

Inputs:

• Y: image (spatial dimension 1 x spatial dimension 2 x spectral dimension=d).
• R: number of endmembers.
• K: number of clusters.
• C: number of classes.
• labels: label matrix where label 0 corresponds to unknown class to be estimated.
• maxIter: maximum number of iterations of PALM algorithm.
• classif_loss: loss function to use for classification term ('quadratic'|'cross-entropy')
• M: endmembers matrix with R spectra (dxR).
• maxInitIter: maximum number of iterations for algorithm initializing A.
• adaptTV: (bool) use TV weighted by the gradient of the penchromatic image.
• printEnable: (bool) print information when True. Default False.
• initClassif: (bool) use random forest to initialize classes.
• initFilePath: (bool) path to store results of initialization, (if maxInitIter=0, path to load initialization)
• srand: random seed.

Ouput:

• A_est: estimated abundance matrix.
• M: endmember matrix.
• B_est: estimated cluster centroid matrix.
• Z_est: estimated cluster matrix.
• Q_est: estimated interaction matrix.
• C_est: estimated classification matrix.
• objectiveFct_tab: historic of the value of the terms of the objective function.
• timeInit: processing time due to initialization.

Authors

Author: Adrien Lagrange (ad.lagrange@gmail.com)

Under Apache 2.0 license