1. Introduction

Training Artificial Intelligence (AI) models to generate/predict new built environment objects (e.g. roads, buildings, etc.) directly from remote sensing imagery shows great promise for automating Volunteered Geographic Information (VGI) mapping workflows [1]. However, currently there is a dis-inclination to accept such AI generated map features into VGI platforms like OpenStreetMap (OSM) due to concerns over their geometric quality. In particular, existing approaches to assess building footprint (outline) quality/accuracy typically evaluates for positional accuracy or data completeness - but lacks a standardized methodology for measuring their shape complexity [2]. This paper addresses this concern by proposing a novel Shape Quality Metric (SQM) designed specifically to evaluate the geometric fidelity of AI-generated building outlines. SQM provides a robust, systematic metric to quantify building shape quality, allowing for reliable comparison/integration of realistic AI-derived map features into the online OSM database. Our approach trains GeoAI models using readily available Google Earth imagery and OSM vectors and validates them against authoritative "ground truth" datasets from Ordnance Survey Ireland (OSi) [3], [4]. Through a series of quantitative assessments, we show how SQM effectively addresses key VGI mapping concerns, such as geometric accuracy and completeness, and introduces a systematic measure to evaluate overall reliability of AI-generated map data. This work contributes to automating the mapping of remotely sensed buildings by bridging the gap between AI and VGI, paving the way for more accurate, up-to-date, and reliable online maps.

2. Aim of the Study

The primary aim of this study is to introduce and validate a reliable, systematic measure of geometric fidelity for GeoAI-generated building footprints. By providing a robust framework to quantify shape complexity independent of positional accuracy and completeness, the proposed Shape Quality Metric aims to bridge the gap between AI-driven mapping and crowdsource mapping platforms. The ultimate goal of SQM is to increase trust and confidence among the VGI mapping community in adopting AI-generated map features into the live OSM database.

3. Methodology

The Shape Quality Metric is proposed as a robust framework to systematically evaluate the geometric fidelity of building outlines (footprints), specifically those predicted/generated by GeoAI models. Following established GIS criteria, the methodology employed six key polygonal parameters: Area, Perimeter, Elongation, Convexity, Compactness, and Solidity [5], [6], [7], [8]. These fundamental shape descriptors were selected based on a correlation analysis of 3,036 building objects from authoritative OSi "ground truth" data to ensure each provides unique, non-redundant geometric information.
Before integration, each shape descriptor is normalized to a [0, 1] range using min-max normalization, where a value of 1 represents an ideal geometric form. SQM is then calculated as a non-linear product ratio designed to amplify poor geometric characteristics like jaggedness and fragmentation:
SQM=Perimeter*Elongation*Convexity/Area*Compactness*Solidity
This formula balances Complexity Drivers in the numerator (e.g., perimeter, which increases with boundary instability) against Quality Stabilizers in the denominator (e.g., solidity, which rewards well-filled, convex footprints). The resulting metric is dimensionless and intentionally unbounded, allowing it to be applied to individual building objects independently without requiring dataset-wide normalization.
To validate the metric, experiments were conducted within the DeepMapper mapping platform [9]. Building footprints were first generated using two separate Generative Adversarial Network (GAN) models:

1. OSi-GAN, trained on authoritative OSi aerial orthophotos (25cm/pixel).
2. OSM-GAN, trained on OSM vectors and Google Earth satellite imagery (30cm/pixel).

A subsequent post-processing phase then applied the Poly-GAN simplification/smoothing algorithm to regularize the outlines/edges of these AI-generated buildings. The final GeoAI-derived buildings were then compared against live OSM data and OSi reference buildings using a one-to-one spatial matching procedure. All buildings were projected to the Irish Transverse Mercator (EPSG:2157) system to ensure a consistent metric evaluation and matched based on the highest Intersection-over-Union (IoU) evaluations. This matching process allowed for a precise pairwise SQM analysis to determine how closely AI-generated shapes mimic both community-mapping (e.g., OSM) and authoritative-mapping (e.g., OSi) standards.

4. Findings

Experimental results comparing 453 conjugate buildings in Dublin city centre of OSi, OSM, and GeoAI datasets demonstrates that SQM effectively captures the key shape-complexity characteristics:

1. Sensitivity: SQM reliably distinguished between simple and complex geometries. Lower SQM values corresponded to higher geometric quality (regular, convex footprints), while higher values flagged lower quality building outlines (fragmented or jagged footprints).
2. Dataset Comparison: Reference OSi buildings exhibited the highest average complexity (Mean: 6.10), reflecting detailed manual surveying methods. Conversely, OSM buildings showed reduced complexity (Mean: 5.49), likely due to variable community-mapping generalizations of complex shapes.
3. AI Performance: AI-generated buildings (OSi-GAN Mean: 6.54; OSM-GAN Mean: 5.32) aligned closely with their respective training sources. Interestingly, AI-generated buildings often showed a promising alignment with OSi reference shapes, often outperforming existing OSM footprints in geometric consistency.
4. Robustness: Controlled perturbations (vertex displacement and boundary simplification) confirmed that SQM behaves predictably, increasing when outlines become more irregular.

5. Discussion & Conclusion

The introduction of SQM provides a scalable, objective layer of analysis for OSM quality assurance that complements traditional manual inspection. By isolating shape fidelity, the metric enables the scientific community and VGI mappers to systematically identify "AI slop" or geometric discrepancies before they are committed to the online database.
Scientific Contribution: The research contributes a novel, dimensionless, and unbounded building shape quality metric that can be applied to individual building objects without the need for dataset-wide normalization. This makes it uniquely suited for real-world, automated map-updating workflows where new features are processed independently.
Practical Benefits & Implications: For the OSM community, SQM offers a tool for initial screening and benchmarking of GeoAI data. It establishes a semi-automated workflow where VGI contributors focus their time on validation rather than manual digitization. By strengthening trust in AI-generated data, this work paves the way for more accurate, up-to-date, and reliable online maps. Future work will explore extending SQM to other VGI map features, such as transport infrastructure (roads, rails, bike paths, etc.) and land-use features (farmland, parkland, industrial, residential, etc.) to further enhance the legitimacy of GeoAI-assisted mapping.