GROVE: A RAG-Enhanced Local LLM Framework for Scalable Urban Forest Carbon Storage Estimation

Jiawei Tong, John Moraros, Shuihua Wang, Guangyu Wang, Sirui Li, Ruoxi Liao

Keywords: Carbon Storage Estimation, Environmental Monitoring, Retrieval-Augmented Generation, Explainable Environmental AI, Multimodal AI

TL;DR: GROVE combines AI vision, scientific knowledge retrieval, and local llm reasoning to estimate forest carbon storage with expert-level accuracy at much lower cost, enabling scalable climate monitoring for environmental policy.

Abstract:

Accurate estimation of forest carbon storage is critical for climate mitigation and environmental policy, yet current methods are constrained by a general trade-off between accuracy and scalability. Traditional expert assessments achieve
high precision but incur prohibitive costs, while automated
vision-based systems sacrifice accuracy for efficiency, creating fundamental barriers to large-scale carbon monitoring.
These barriers include: (1) the high cost of field measurements limiting assessment scope, (2) poor integration of heterogeneous data sources (e.g., degraded imagery, noisy measurements, and environmental context) that reduces prediction reliability, and (3) a lack of scientifically-grounded explainability that undermines policy adoption. To bridge this
gap, we introduce GROVE (Grounded Retrieval Optimized
Vision Estimation), a novel framework that integrates three
core stages: 1) botanical image enhancement to transform degraded field imagery into analysis-ready representations; 2)
retrieval-augmented generation (RAG) to ground predictions
in validated scientific literature and species databases; and
3) hierarchical reasoning via a locally-deployed 7B parameter language model to synthesize multimodal information
into carbon estimates with explicit uncertainty quantification.
Validated on 25,000 images spanning 50 species across diverse global regions, GROVE achieves accuracy approaching expert-level performance (standardized RMSE: 0.42 vs.
0.39 for the i-Tree baseline) while substantially reducing operational costs and enabling offline deployment. Our work
demonstrates that the principled integration of visual data, retrieved scientific knowledge, and structured reasoning can simultaneously deliver scalability and scientific credibility, offering a viable pathway for evidence-based climate policy.

Submission Number: 17