Constraint Based Evaluation of Generalized Images Generated by Deep Learning

The use of deep learning techniques for map generalisation raises new problems regarding the evaluation of the results: (1) images are used as input/output instead of vector data; (2) the deep learning processes do not guarantee results that follow cartographic principles; (3) the deep learning models are black boxes that hide the causal mechanisms. Also, deep learning intern evaluation is mostly based on the realism of the images and the pixel classification accuracy, and none of these criteria is sufficient to evaluate a generalisation process. In this article, we propose an adaptation of the constraint-based evaluation to the images generated by deep learning. Six raster-based constraints are proposed for a mountain road generalisation use case.