A concrete contribution to cadastral maintenance
The Geneva case study combines a Mask R-CNN detector with existing aerial imagery and cadastral records. The model produces object masks, which are transformed back into geographic coordinates and compared with the register. Curating inconsistent reference data raised the reported F1 score from approximately 82% to approximately 90%.
Expert review established the practical outcome: among 562 potentially new objects, 223 were swimming pools subject to registration. Another 211 were pools that did not require registration, and 128 were other objects. Automated detection therefore provides concrete leads for inspection. Deciding whether a detection belongs in the cadastre remains a domain decision.
Evidence [1]
Transfer to another canton
For Thurgau, reference annotations from Geneva and Neuchâtel were combined with SWISSIMAGE aerial imagery. This enabled the detector to be applied to another canton without first digitising a complete new local training inventory. The workflow produced GIS-compatible object outlines and examined the effects of image resolution and confidence thresholds.
The report also shows why a register is not an infallible benchmark: image acquisition dates, subsequently built structures and the distinction between swimming pools and other water basins affect the comparison. Reproducible processing and documented domain review are therefore as important as detection metrics.
Evidence [2]
Methods in detail
- Detect image-based objects with deep learning and convert masks to GIS-compatible outlines.
- Transfer models and reference data between cantons and submit detections to expert review.
- Investigate separate LiDAR processing and segmentation methods for tree detection.
Publications and sources
- Technical reportCerioni & Meyer (2021) · Geneva pool detection
- Technical reportMeyer & Cerioni (2021) · Thurgau pool detection







