EcoScapes: LLM-Powered Advice for Crafting Sustainable Cities

Climate adaptation is critical for the functionality and quality of life in urban areas under more frequent and severe extreme weather events, such as heatwaves, droughts, and floods. Smaller towns, however, may struggle to adapt because of funding issues, administrative burdens and difficulties using environmental data. This study presents EcoScapes, a decision-support framework to enhance LLM advisory with local Earth observation data. EcoScapes integrates three key components: automated acquisition and preprocessing of Sentinel-2 imagery; Vision Language Models (VLMs) for structured interpretation of satellite-derived representations; and a downstream knowledge-based advisory workflow inspired by prior work.

Given a user-provided town or city name, EcoScapes geocodes the location and retrieves Sentinel-2 imagery for a 5 km bounding box around the urban center. To minimize cloud interference, we use a 1% cloud cover filter, which enables usability but might bias towards drier conditions and miss seasonal water bodies. EcoScapes processes satellite data rendered by the Sentinel-2 API, which includes RGB, water, and moisture views. The system uses a modular analytical pipeline, with an RGB analysis module employing a VLM to describe urban structures, like built-up areas, green spaces, and roads, via focused prompts. This approach reduces hallucinations and ensures more accurate analyses. Separate water and moisture modules analyze the outputs. Water analysis removes small, likely irrelevant features before an RGB-based step connects identified water bodies to their environment and infrastructure. Moisture analysis is used to find heat islands. Finally, a local small language model combines outputs into a single “Climate Report”. This report is subsequently used as context for a ChatClimate-style [1] system that is grounded in the IPCC AR6. This enables a comparison between a baseline advisory system relying on the knowledge base alone and the same system augmented with EcoScapes’ local report.

Since EcoScapes generates varied text outputs, we qualitatively assess its performance using two contrasting case studies: Roßtal, a small rural community of 10,000 people, and Erlangen, a medium-sized city with a population exceeding 100,000. The results indicate EcoScapes can provide useful local context where pre-existing model knowledge is limited. EcoScapes’ report made Roßtal’s adaptation recommendations more relevant and usable, correcting geographically inaccurate suggestions in the baseline. However, EcoScapes’ own inconsistencies and occasional hallucinations remain a limiting factor. The downstream recommendations were affected by errors in interpreting water data in Erlangen, relative to the baseline system, which was more familiar with the city because of its training data. EcoScapes demonstrates Sentinel-2 data’s potential to improve climate advice in smaller towns. Achieving generalization will require improved multimodal reasoning and higher resolution images, while broader evaluation is necessary to determine whether such generalization holds.

More information can be found at our GitHub repository (https://github.com/Photon-GitHub/EcoScapes) and the corresponding paper on arXiv (https://arxiv.org/abs/2512.14373).

 

References

[1] S. Vaghefi et al., “Chatclimate: Grounding conversational ai in climate science,” Communications Earth & Environment, vol. 4, no. 1, pp. 480, 2023