Våren 2025 har det vært enorm aktivitet i KartAi-prosjektet! Et skred av akademiske prosjekter med tilhørende rapporter har blitt gjennomført, produsert og evaluert. Samtlige har levert svært høyt nivå.
Du finner alle rapportene under Rapporter og Resultater
I god KI-ånd har vi oppsummert alle rapportene med NotebookLM. Hør deep dive podcast her:
Den mer tradisjonelle skriftlige oppsummeringen finner du under:
- GeoGPT (IS-304, Group 1): This bachelor project, conducted by the GeoGutta student group at the University of Agder, aimed to further develop GeoGPT, building on a previous proof-of-concept (PoC) to deliver a minimum viable product (MVP). The goal was to demonstrate how large language models (LLMs) can make map data more accessible and understandable for end-users. The project, undertaken in collaboration with Kristiansand municipality and Kartverket, was part of Kartverket’s broader KartAi project. The methodology employed was Agile, specifically Scrum, with principles from Lean Startup also incorporated for rapid hypothesis testing and user involvement. Key developments included implementing a new map solution based on Leaflet, transitioning from LangChain to LangGraph for improved conversational flow, expanding the dataset from 200 to 750 datasets, and migrating to a multi-agent chat architecture for handling complex user inquiries and map interactions. User testing, particularly at Geomatikkdagene, led to a significant redesign, consolidating the chatbot and map catalog into a side panel for better user experience. The final product (DemoV7) focused on improved user interface and increased functionality, offering both standard Geonorge search and GeoGPT AI search in the map catalog.
- GeoLake (IS-304): This bachelor thesis, completed by the TechTroll student group at the University of Agder, focused on how modern streaming technologies can be used for efficient sharing of geodata, aiming to reduce download times and improve user experience. The project was a part of the KartAI project and was conducted in collaboration with UiA, Norkart, Kartverket, and Tietoevry from January to May 2025. The project used Python and Jupyter notebooks for development. It explored and compared three solution models for processing and visualizing geographical data: GeoPandas, DuckDB, and Databricks, based on AR50- and AIS-data. The methodology used was Agile, with a strong emphasis on Scrum, including daily Scrums, sprint planning, and reviews/retrospectives. The project also developed a «roadmap» and «capacity plan» for quality assurance and time management. The results demonstrated that streaming of geodata has significant potential to replace manual download processes, providing interactive notebooks tailored to specific user stories. The group recommended Databricks for future MVP development due to its strong support for Delta Lake and Delta Sharing.
- Digital Transformation in the Norwegian Public Sector: A Case Study of Fellestjenester BYGG (IS-305, Group 13): This research assignment by Group 13 from the University of Agder examined how digital transformation impacts efficiency and service delivery in the Norwegian public sector, using Fellestjenester BYGG as a case study. The project was inspired by a collaboration with Kristiansand Municipality. The methodology involved a descriptive qualitative case study approach with a literature review, policy analysis, and qualitative interviews with a senior official from DiBK, a municipal building caseworker, and an architect. Findings indicated that Fellestjenester BYGG improves front-end processes for external users (e.g., architects) through automation, but internal users face challenges like limited onboarding, poor system integration with legacy workflows, and lack of co-creation. The study concluded that successful digital transformation requires inclusive design, structured onboarding, and ongoing user engagement, emphasizing that efficiency gains are undermined by misaligned internal processes and fragmented municipal capacity due to Norway’s decentralized governance.
- BirdAI: This bachelor project was a collaboration between a student group (Jonas Fritzøe Hovdenak, Sander Javier Nomedal, Martin Steiro, Ruben Teikari, Johannes Tjøstheim), Kartverket (geodesy department), and Tietoevry. The primary goal was to develop an MVP model for real-time detection of relevant objects, with a specific real-world application for safety around SLR lasers. A major challenge faced was the initial minimal domain knowledge of machine learning. The project utilized various object detection models such as RT-DETR, Yolo (v11, v9), RF-DETR, and most notably D-FINE, trained using Google Colab. The dataset, comprising approximately 8000 images, was developed using Roboflow for annotation and augmentation. The group expressed satisfaction with the learning experience in machine learning, datasets, development, and project management, highlighting the valuable guidance and resources from Kartverket and Tietoevry.
- AI Segmentation and Quality Control of Buildings (Master’s Thesis by Marianne Andersen): This master’s thesis, part of the KartAi initiative at NTNU, explored the use of deep learning methods (YOLOv8-seg and Mask R-CNN) to support Kartverket’s completeness check in mottakskontroll for FKB building data. The objective was to evaluate if these models could reliably detect missing or excess buildings in aerial imagery to reduce manual inspection workload. The study used aerial images and FKB building data from Farsund municipality in Norway. YOLOv8-seg showed higher overall precision and fewer false positives than Mask R-CNN, while Mask R-CNN achieved higher recall but more false detections. Post-processing with the SAHI framework improved detection consistency, especially at tile edges. The conclusion was that while neither model could fully replace manual control, YOLOv8-seg showed strong potential as a decision-support tool to guide attention to areas likely to have discrepancies, thereby increasing inspection efficiency.
- AI-based Approaches for Quality Control of Map Data (Master’s Thesis by Jakob Severin Steffensen Hjelseth): This master’s thesis, a cooperation between Kartverket, Norkart, and NTNU in the KartAI collaboration, investigated the use of deep learning networks for semantic segmentation of roof edges (buildings) and road edges (asphalt) for quality control of map data. The goal was to determine if these networks could achieve sufficient locational accuracy to automate the geographic comparison in quality control. The project focused on Farsund, Norway, a varied geographic area, for data. The chosen network was FarSeg, applied with PyTorch and TorchGeo, and the methodology involved pre-processing data into tiles, training the model, performing segmentations, and validating results. Results showed strong performance comparable to or better than state-of-the-art techniques on Norwegian data, indicating that DL networks can be effectively adapted to diverse terrain characteristics. The thesis suggests that while full automation is not yet feasible, AI-assisted tools could significantly reduce manual workload in map control by identifying specific features.
- PlanAID – System Development (Bachelor’s Thesis by Ole Sveinung Berget and Are Berntsen): This bachelor’s thesis, conducted at the University of Agder in collaboration with Evje and Hornnes municipality, aimed to make zonal planning more efficient and user-friendly. It focused on creating a prototype system that could analyze, compare, and visualize how regulations across different zonal plans interact, primarily by automating the extraction, comparison, and visualization of planning documents. The project integrated rule-based processing with strategic AI integration (e.g., NER for field extraction), complementing NLP components from a companion thesis. PlanAID’s core technical contributions include an effective document processing pipeline for Norwegian planning documents, a hybrid approach to terminology normalization, a modular microservice-based system architecture for planning analytics, and novel visualization techniques for cross-document regulatory comparison. The system processes PDF regulations, PDF maps, and SOSI files. Key findings indicated that the microservice architecture proved highly effective, and Docling-based hierarchical structuring was fundamental for accurate and structured outputs. The system reduced manual comparison time from days to typically under 30 seconds for initial analysis.
- Applying Transformer-Based Language Models to Automate Named Entity Recognition and Summarization in Municipal Zoning Plans (Bachelor’s Thesis by Simen Bihaug-Frøyland): This bachelor’s thesis explored the use of transformer-based language models (BERT and GPT) to automate two key tasks in municipal zoning plans: recognizing field zone names (Named Entity Recognition – NER) and summarizing changes between plan revisions. The project was part of the PlanAId initiative and used real-world data from Arealplaner.no. For NER, fine-tuned BERT models achieved high F1-scores (e.g., nb-bert-large at 91.23%), outperforming prompt-based LLMs. For change summarization, LLMs (specifically GPT-4) generated coherent and accurate summaries, achieving a ROUGE-L F1-score of 63.41%. The study found that fine-tuned, domain-adapted transformer models are best for accurate entity extraction in specialized documents, while autoregressive LLMs are better for creating clear, high-quality summaries. The use of Genetic Algorithms (GAs) to optimize prompt examples led to measurable performance improvements. The project created a new annotated dataset for Norwegian zoning plan texts.
- TiltaksAid (IS-304): This bachelor project by a student group at Kristiansand Municipality aimed to further develop a previous prototype for «Digitale tiltak», focusing on making the building application process less complex and more user-friendly through visualization and adherence to regulations. The project was conducted within the IS-304 course. The group used an Agile methodology inspired by Scrum, with the project divided into five sprints. Trello was used for project management, and Google Docs for meeting minutes. Key technologies included Figma for UI/UX design, QGIS and PostGIS for geodata processing, Leaflet and OpenStreetMap for interactive map visualization, and Supabase for backend and database management. The system incorporated AI predictions of unregistered buildings and developed an automated checklist for building permit criteria (e.g., minimum distances to property boundaries, existing buildings, and roads) using PostGIS for spatial calculations. A significant learning was that AI cannot operate without human validation («human-in-the-loop») in systems affecting individuals.
- Utstillingsvindu 2.0 – Practical Application of KartAi Technology (IS-304): This bachelor project from the University of Agder aimed to streamline the building application process by integrating AI agents into a functional prototype, focusing on reducing complexity and improving submission quality for one-step building applications. It built upon a previous Proof of Concept within the broader KartAi initiative, in collaboration with Kristiansand Municipality. The group used the agile Scrum methodology with 17 sprints, emphasizing a user-centered approach for applicants. Key technologies used included the T3 Stack (React, Next.js, TypeScript, TailwindCSS, tRPC, Prisma, MySQL) for full-stack development, and Python. It integrated specific KartAi APIs: TiltaksAiD for spatial zoning analysis and CADAiD for document validation, deliberately limiting the number of AI technologies to avoid confusing users. The project aimed to deliver a high-fidelity prototype as a foundation for further development. AI tools like GitHub Copilot and ChatGPT were used for coding assistance and linguistic quality of the report, with careful verification of AI-generated outputs. The final prototype supports the «before» and «during» phases of a building application, with partial implementation of the «after» phase.
