Strategic Afforestation Planning in Denmark - A Multi-Criteria Framework Integrating Large Language Models for AHP-Based Expert Elicitation
Author
Elmegaard-Fessel, Christina Louise
Term
4. term
Publication year
2025
Submitted on
2025-05-28
Pages
66
Abstract
Dette speciale tager udgangspunkt i den grønne trepartsaftale, som kræver skovrejsning på 250.000 ha i Danmark inden 2045 for at styrke CO₂-binding og biodiversitet. Formålet er 1) at udvikle en rumligt eksplicit ramme, der prioriterer skovrejsningsområder på tværs af miljømål, landbrugsinteresser og lokal planlægning, og 2) at afprøve Large Language Models (LLM’er) som et nyt redskab til ekspertinddragelse i Analytic Hierarchy Process (AHP). Der gennemføres en GIS-baseret Multi-Criteria Decision Analysis med vægtet overlay-analyse i 100 m opløsning for hele landet. Ni rumlige kriterier, udledt af aftalen og MCDA-principper, vægtes med AHP, hvor en panelvurdering fra fem menneskelige eksperter sammenlignes med tre LLM’er (ChatGPT‑o3, Gemini 2.5 Pro, Grok 3), der emulerer forskellige interessentroller; samme setup bruges til at normere kategorier i hvert kriterielag. Juridiske og fysiske begrænsninger anvendes som binær maske for at frasortere uegnede arealer. LLM’ernes samlede vægtvektor viser stærk rangordensoverensstemmelse med den menneskelige (Spearmans rho = 0,69), og de fleste LLM‑matricer har acceptabel konsistens, hvilket peger på, at LLM’er kan levere plausible, politikrelevante AHP‑vurderinger som et tids- og omkostningseffektivt supplement – men ikke en erstatning – for menneskelig ekspertise. Suitabilitetsanalysen identificerer 1,873 mio. ha egnede arealer efter begrænsninger, heraf 1,058 mio. ha som høj prioritet (score 8–9). En prioriteret portefølje på 250.000 ha udpeges primært fra højtscorende områder (9, 8 og nødvendigvis 7) i tilknytning til eksisterende skov for at styrke landskabssammenhæng; lokal kvalitativ validering underbygger modellens rumlige logik. Specialet konkluderer, at den udviklede GIS‑baserede MCDA‑ramme er gennemsigtig og tilpasningsdygtig til strategisk skovrejsningsplanlægning, og at AHP understøttet af LLM’er er en levedygtig, effektiv metode til ekspertinddragelse, forudsat solid menneskelig faglighed i promptdesign, validering og fortolkning. Studiet leverer en rumligt eksplicit portefølje til at understøtte implementeringen i Danmark og peger samtidig på metodiske begrænsninger, skiftende politiske vilkår og praktiske udfordringer som eksisterende tilskudsordninger.
This thesis addresses Denmark’s Green Tripartite Agreement, which mandates 250,000 hectares of afforestation by 2045 to enhance CO₂ sequestration and biodiversity. It pursues two aims: (1) to develop a spatially explicit framework that prioritizes afforestation areas while balancing environmental goals with agricultural and local planning considerations, and (2) to assess Large Language Models (LLMs) as a novel tool for expert elicitation within the Analytic Hierarchy Process (AHP). A GIS-based Multi-Criteria Decision Analysis was implemented using a weighted overlay at 100 m resolution nationwide. Nine spatial criteria, derived from the agreement and MCDA principles, were weighted with AHP by benchmarking a five-person human expert panel against three LLMs (ChatGPT‑o3, Gemini 2.5 Pro, Grok 3) emulating stakeholder roles; the same setup was used to generate category normalization scores. Legal and physical constraints were applied via a binary mask to exclude unsuitable areas. For expert elicitation, the LLM composite weight vector showed strong rank-order agreement with the human-derived vector (Spearman’s ρ = 0.69), and most LLM matrices exhibited acceptable consistency, indicating that LLMs can produce plausible, policy-relevant AHP judgments efficiently as a complement—though not a substitute—to human expertise. The suitability analysis identified 1.873 million hectares of land suitable after constraints, including 1.058 million hectares as high priority (scores 8–9). A 250,000‑hectare prioritized portfolio was delineated primarily from high-suitability areas (scores 9, 8, and necessarily 7) adjacent to existing forests to strengthen landscape coherence, and local qualitative checks supported the model’s spatial logic. The thesis concludes that the GIS‑based MCDA framework is transparent and adaptable for strategic afforestation planning, and that LLM‑assisted AHP is a viable, efficient method for expert elicitation when paired with strong human oversight in prompt design, validation, and interpretation. It delivers a spatially explicit portfolio to inform Danish policy implementation while acknowledging methodological limitations, evolving policy contexts, and practical issues such as existing subsidy schemes.
[This abstract was generated with the help of AI]
Keywords
Documents