Semantic Segmentation of Buildings Using Multisource ALS Data

In this paper we propose to utilize a SparseCNN network and national ALS datasets of Vienna and Zurich to achieve generalization of a classifier by including both datasets simultaneously in the training phase. The data was classified into ground and water, vegetation, building and bridges, and 'other'. The results were evaluated using median IoU. The classifier trained with both datasets performed only slightly worse on ground and water and on vegetation in comparison to the classifiers trained and tested using dataset from the same city (maximum drop of 0.3 pp from a value above 94%). For building and bridges the accuracy slightly improves (at least 0.6 pp), whereas for 'other' results are inconsistent. The classifier trained using both datasets performed substantially better than the classifiers trained using one dataset and tested on the other. Thus, training using multiple datasets leads to a more general classifier while maintaining accuracy.