EcoScapes: LLM-Powered Advice for Crafting Sustainable Cities

Climate adaptation is vital for the sustainability and sometimes the mere survival of our urban areas. However, small cities often struggle with limited personnel resources and integrating vast amounts of data from multiple sources for a comprehensive analysis. To overcome these challenges, this paper proposes a multi-layered system combining specialized LLMs, satellite imagery analysis and a knowledge base to aid in developing effective climate adaptation strategies. The corresponding code can be found at https://github.com/Photon-GitHub/EcoScapes.

💡 Research Summary

EcoScapes presents a proof‑of‑concept decision‑support framework that aims to empower small‑to‑medium sized cities with automated, locally‑tailored climate‑adaptation advice. The authors begin by highlighting the disproportionate challenges faced by resource‑constrained municipalities—limited staff, tight budgets, and the difficulty of integrating heterogeneous data sources. While recent AI advances, especially large language models (LLMs), have shown promise in simplifying climate‑adaptation planning (e.g., the ChatClimate system that grounds GPT‑4 outputs in IPCC AR6), these systems still lack geographic specificity.

To fill this gap, EcoScapes combines multimodal LLMs with Sentinel‑2 satellite imagery. The data acquisition pipeline uses SentinelHub (with a fallback to manual Copernicus downloads) to retrieve RGB, near‑infrared, moisture, and water‑index images for a 5 km bounding box around a user‑specified town. A strict cloud‑cover filter (<1 % cloud) reduces atmospheric interference but introduces a bias toward drier conditions. The authors list the relevant Sentinel‑2 bands (B02, B03, B04, B08, B8A, B11) and compute moisture and water indices via a normalized difference formula.

The system architecture is modular. A “Satellite Loader” converts place names to coordinates via Nominatim, fetches the images, and stores them as rgb.png, moisture.png, and water.png. Each subsequent analytical module inherits from a generic Module class that tracks hard and soft dependencies, allowing non‑critical failures without halting the whole pipeline. The core image‑analysis modules employ the 360VL model—a perception‑oriented multimodal LLM—to interpret the images. RGB analysis identifies built structures, green spaces, and road patterns; the water pipeline pre‑processes the water index (thresholding, denoising) before re‑examining the RGB image for proximity to infrastructure; the moisture module flags anomalies linked to urban heat islands or cooler zones.

For report generation, the authors select InternLM, an open‑source model under 10 B parameters, suitable for local deployment. A system prompt instructs InternLM to act as a climate scientist, ensuring that the generated text is grounded in scientific reasoning. The resulting climate‑adaptation report is then fed into a recreated ChatClimate (CC) knowledge base, which also contains the IPCC AR6 as a reference file. This integration allows CC to augment its generic advice with the locally‑derived insights from EcoScapes.

Evaluation is qualitative, reflecting the difficulty of benchmarking creative, context‑aware outputs. Two contrasting case studies are examined: Roßtal (≈10 k inhabitants, limited resources) and Erlangen (≈100 k inhabitants, more complex infrastructure). The authors devise scoring rubrics for EcoScapes reports (Correctness and Depth & Coverage, each 0‑5) and for the final CC strategies (Usability, Correctness, Relevance, each 0‑5). Manual graders assign scores based on predefined criteria, with attempts to blind the source of the texts, though complete blinding proved impractical due to stylistic cues.

Results show that EcoScapes struggles with correctness, especially in heat‑island analysis where hallucinations produced false features. Depth and coverage scores are higher, indicating that the system can reliably extract basic land‑cover information. When integrated with CC, the locally‑specific reports modestly improve relevance scores, but overall usability and factual accuracy of the final recommendations remain limited. The authors attribute these shortcomings to several sources of bias: (1) environmental bias from satellite limitations (clouds, seasonal variation, resolution); (2) policy bias embedded in the pre‑training data of the LLMs; (3) model bias inherent to LLM hallucinations; and (4) evaluator bias despite attempts at standardization.

The paper concludes that EcoScapes demonstrates a viable architecture for augmenting climate‑adaptation planning in small cities, but significant technical hurdles remain. Future work should explore higher‑resolution commercial satellite data (e.g., PlanetScope), incorporate fact‑checking or retrieval‑augmented generation to curb hallucinations, expand testing across diverse climatic zones and city sizes, and develop more objective, possibly automated, evaluation metrics. The authors provide all code and full report texts on GitHub, encouraging community replication and extension.