[Paper] NordFKB: a fine-grained benchmark dataset for geospatial AI in Norway

Source: arXiv - 2512.09913v1

Overview

A new open‑source dataset called NordFKB brings high‑resolution aerial imagery and meticulously curated annotations to the geospatial AI community—specifically for Norway. By pairing orthophotos with 36 fine‑grained semantic classes and both segmentation masks and bounding‑box labels, the authors aim to accelerate research and real‑world applications in mapping, land‑use analysis, and spatial planning.

Key Contributions

• Fine‑grained benchmark: 36 semantic classes (e.g., roads, water bodies, building types) with per‑pixel masks and COCO‑style bounding boxes.
• High‑quality source data: Built on Norway’s authoritative Felles KartdataBase (FKB), guaranteeing sub‑meter positional accuracy.
• Geographically diverse sampling: Tiles drawn from seven regions covering different climates, topographies, and urbanization levels.
• Balanced, representative splits: Randomized training/validation splits that preserve class distribution across all areas.
• Reproducible evaluation suite: Open‑source repository with standardized metrics, scripts, and baseline models for semantic segmentation and object detection.
• Human‑in‑the‑loop QC: Expert review of every annotation to ensure consistency and correctness.

Methodology

1. Data acquisition – The team extracted orthophoto tiles (≈0.5 m resolution) from the national FKB repository. Only tiles containing at least one annotated object were kept to maximize label density.
2. Class definition & annotation – 36 classes were defined in collaboration with domain experts (urban planners, cartographers). Annotators produced both binary masks (GeoTIFF) and bounding boxes (COCO JSON).
3. Quality control – After initial labeling, a second expert reviewed each tile, correcting errors and harmonizing class boundaries.
4. Split generation – Tiles were randomly sampled across the seven regions, producing training and validation sets that reflect the full geographic and class variability.
5. Benchmarking toolkit – The authors packaged the dataset with Python utilities (PyTorch‑compatible dataloaders, evaluation scripts) and baseline models (U‑Net for segmentation, Faster‑RCNN for detection) to lower the entry barrier for new researchers.

Results & Findings

• Baseline performance: Using a standard U‑Net, the authors achieved a mean Intersection‑over‑Union (mIoU) of 68.4 % across all 36 classes; Faster‑RCNN reached a mean Average Precision (mAP) of 57.1 % for object detection.
• Class imbalance impact: Rare classes (e.g., “railway bridge”) showed significantly lower scores, highlighting the need for advanced sampling or loss‑balancing techniques.
• Geographic transferability: Models trained on three regions generalized reasonably well to the remaining four, but performance dropped by ~5 % on the most topographically extreme area, suggesting that terrain diversity still challenges current architectures.
• Annotation fidelity: Human QC reduced label noise to an estimated <1 % error rate, confirmed by a spot‑check audit of 500 random objects.

Practical Implications

• Rapid map updating: Developers can fine‑tune segmentation models on NordFKB to automatically extract building footprints, road networks, or water bodies from new aerial surveys, cutting manual cartography time dramatically.
• Smart city & infrastructure planning: High‑resolution object detection enables automated inventory of assets (e.g., streetlights, parking lots) for municipal asset management systems.
• Environmental monitoring: Precise land‑cover masks support change‑detection pipelines for flood risk assessment, deforestation tracking, and biodiversity studies.
• Commercial GIS services: Companies building location‑based services can leverage the dataset to train domain‑specific models that outperform generic off‑the‑shelf solutions, leading to better product differentiation.
• Education & research: The open benchmark lowers the barrier for university labs and hobbyist developers to experiment with state‑of‑the‑art geospatial AI techniques without costly data licensing.

Limitations & Future Work

• Temporal staticity: All images are from a single acquisition period; seasonal or multi‑temporal analysis is not possible yet.
• Geographic scope: Although diverse, the dataset covers only seven regions, leaving large parts of Norway under‑represented.
• Modalities: Currently limited to RGB orthophotos; adding LiDAR, multispectral, or SAR data would broaden applicability.
• Class granularity vs. rarity: Some fine‑grained classes have very few instances, which hampers robust model training—future releases may merge or augment these categories.

NordFKB opens the door for a new wave of geospatial AI applications in Norway and beyond. By providing both the data and the tooling, the authors set a solid foundation for reproducible research and real‑world impact.

Authors

• Sander Riisøen Jyhne
• Aditya Gupta
• Ben Worsley
• Marianne Andersen
• Ivar Oveland
• Alexander Salveson Nossum

Paper Information

• arXiv ID: 2512.09913v1
• Categories: cs.CV
• Published: December 10, 2025
• PDF: Download PDF