Geospatial AI-Agent Workflows for Land Management Applications
Author
Adiyasa, Alexander
Term
4. semester
Publication year
2025
Abstract
Denne afhandling undersøger, hvordan naturlig sprogbehandling og geospatial analyse kan understøtte kommunernes GIS-baserede konfliktvurdering i sager om dispensation til arbejde ved vejafvandingsbassiner beskyttet efter naturbeskyttelseslovens §3. Gennem en iterativ metode baseret på interessentinterviews og analyse af tidligere dispensationssager udvikles en geospatial AI-agent, der kombinerer en interaktiv chatbot til rumlige forespørgsler med en ny rapportfunktion. Her omsætter en opgaveinstrueret Large Language Model (LLM), guidet af domænespecifikke instruktioner, deterministiske geoværktøjers output til udvidede konfliktvurderingsrapporter med indledende sagsrelevante konsekvenser. Evalueringen viste høj faktuel nøjagtighed i geospatiale screeninger og 97% succes i LLM’ens anvendelse af betinget logik til at udlede sags-specifikke reguleringsimplikationer. Udfordringer relaterede sig især til datakvalitet og LLM’ens tolkning af rumlige nuancer, der ikke fanges i data. Studiet peger på et betydeligt potentiale for at lette sagsarbejdet og styrke eksisterende praksis, og anbefaler samskabelse af LLM-instruktioner med fageksperter som en central metode for videre udvikling.
This thesis examines how natural language processing and geospatial analysis can enhance the GIS-based conflict assessment used by Danish municipalities when processing dispensations for work at road runoff basins protected under Section 3 of the Nature Protection Act. Using an iterative approach informed by stakeholder interviews and analysis of prior cases, it proposes a geospatial AI-agent that pairs an interactive chatbot for spatial query execution with a novel reporting capability. A task-instructed large language model (LLM), guided by domain-specific instructions, synthesizes deterministic geospatial tool outputs into augmented conflict assessment reports with initial case-specific regulatory implications. Evaluation showed high factual accuracy in geospatial screenings and a 97% success rate in the LLM’s application of conditional logic to derive case-specific implications. Identified challenges include data quality and the LLM’s interpretation of spatial nuances not captured in available datasets. The study indicates substantial potential to support casework and strengthen current practices, and recommends co-designing LLM instruction sets with domain experts as a key strategy for further development.
[This summary has been generated with the help of AI directly from the project (PDF)]
Documents