Datasets collecting demographic and socio-economic statistics are widely available. Still, the data are often only released for highly aggregated geospatial areas, which can mask important local hotspots. When conducting spatial analysis, one often needs to disaggregate the source data, transforming the statistics reported for a set of source zones into values for a set of target zones, with a different geometry and a higher spatial resolution. This article reports on a novel dasymetric disaggregation method that uses encoder–decoder convolutional neural networks, similar to those adopted in image segmentation tasks, to combine different types of ancillary data. Model training constitutes a particular challenge. This is due to the fact that disaggregation tasks are ill-posed and do not entail the direct use of supervision signals in the form of training instances mapping low-resolution to high-resolution counts. We propose to address this problem through self-training. Our method iteratively refines initial estimates produced by disaggregation heuristics and training models with the estimates from previous iterations together with relevant regularization strategies. We conducted experiments related to the disaggregation of different variables collected for Continental Portugal into a raster grid with a resolution of 200 m. Results show that the proposed approach outperforms common alternative methods, including approaches that use other types of regression models to infer the dasymetric weights.