Scalable Matching of Agricultural Field Polygons for Large-Scale Data Comparisons

Advances in deep learning and data harmonization led to a growing availability of large datasets of agricultural field polygons of different granularity and quality. Identifying correspondences between datasets covering the same geographic regions enables opportunities for ●Data Quality Assessment: validation of model outputs against reference ●Data Fusion & Integration: combine complementary information ●Temporal Change Analysis: monitoring of what has changed over time Authors: Sven Gedicke¹, Shiyaza Risvi¹, Alexander Naumann, and Jan-Henrik Haunert¹ (Geoinformation Group, University of Bonn) ¹on behalf of the FAIRagro consortium Scalable Matching of Agricultural Field Polygons for Large-Scale Data Comparisons input overlay correspondences Find matching M = { 𝜇1, … , 𝜇n } that maximizes Matching Approach ●geometry-based matching [2] ●optimization strategy ●graph-based model ●scalability through decomposition connected component matching graph tree graphs Algorithmic Procedure 1. consider connected components separately 2. build hierarchical grouping trees 3. build matching graph 4. solve tree-constrained bipartite matching problem [1] We use IoU, but metric is flexibly adaptable! [1] Canzar, S., Elbassioni, K., Klau, G. W., & Mestre, J. (2015). On tree-constrained matchings and generalizations. Algorithmica, 71(1), 98–119. [2] Naumann, A., Bonerath, A., & Haunert, J.-H. (2025). Scalable many-to-many building footprint matching. Information Fusion, 124, 103360. Visualization Tool Case Study 2 reference vs. reference Case Study 1 model vs. reference