The effect of imposing 'fractional abundance constraints' onto the multilayer perceptron for sub-pixel land cover classification

Stien Heremans, Johan A. K. Suykens, Jos Van Orshoven

Abstract: Highlights•Sub-pixel land cover fractions must be nonnegative and sum to one.•We enforced these constraints onto an MLP, both in-training and post-training.•At the pixel level, both forms of constrained learning markedly improve the results.•At a higher level of spatial aggregation the results are less straightforward.•We recommend in-training constraints for sub-pixel land cover estimation with MLPs.