Developed Region-Specific AI for Climate Change Adaptation Planning

1. Key Points of Announcement

We have developed the world's first large language model (LLM) specialized for regional climate, capable of utilizing regional adaptation plans and ensemble climate projection data to support the planning of climate change adaptation measures (※1).

A case study targeting Kumagaya City, Saitama Prefecture, demonstrated the model's ability to quantitatively propose heat countermeasures (such as the number of green curtains and rest areas to be added) based on future temperature rise projections, along with the calculation process.

This tool is expected to strongly support data-driven, scientific climate change response planning for local governments and small and medium-sized enterprises with limited specialized personnel and resources.

[Glossary]

※1 Climate Change Adaptation Measures

Efforts to mitigate damage to people, society, and ecosystems from current or unavoidable future impacts of climate change.

※2 Ensemble Climate Projection Data

Data that enables probabilistic evaluation of extreme weather events under specific climate conditions by simultaneously conducting numerous climate projection simulations with slightly varied computational conditions.

Reference: https://www.jamstec.go.jp/j/pr/topics/explore-20241223/

※3 Large Language Model (LLM)

A type of AI for natural language processing that can understand and generate human-like natural language by learning from vast amounts of text data from the internet, etc.

2. Overview

Dr. Daisuke Matsuoka, Senior Researcher and Program Director of the Data Science Research Program at the Research Institute for Global Change, Japan Agency for Marine-Earth Science and Technology (JAMSTEC, Secretary-General: Tomohiko Kawamura), in collaboration with Associate Professor Masayuki Hara of the Faculty of Agriculture, Forestry and Marine Science, Kochi University, and Mr. Kazunari Sugiyama, Executive Officer at Ridge-i Inc., has developed a region-specific LLM to support climate change adaptation planning.

Effective adaptation to climate change requires climate information that is scientifically reliable and easy for non-experts to use. This research developed an LLM with expertise in climate science that can also directly search and extract numerical data from future ensemble climate projection data. This method demonstrated excellent performance on a proprietary benchmark for climatology (※4) and, in a proof of concept (PoC) targeting Kumagaya City, Saitama Prefecture, successfully materialized heatstroke countermeasures using probabilistic future temperature rise projections and proposed actionable plans.

This is a pioneering achievement for next-generation climate services, enabling practitioners without advanced specialized knowledge to conduct sophisticated climate risk assessments and develop countermeasures through natural language.

This achievement was published on July 1st (US time) in the Journal of Geophysical Research: Machine Learning and Computation, a journal of the American Geophysical Union. This research was supported by NEDO GENIAC (24036962), the Environment Research and Technology Development Fund (JPMEERF25S12433), the Ministry of Education, Culture, Sports, Science and Technology's "Platform for the Integration and Analysis of Earth Environmental Data" (JPMXD0721453504) and "Advanced Research Program for Climate Change Projection" (JPMXD0722680734), and JSPS KAKENHI (JP22H01316).

[Paper Information]

Title: An LLM Framework for Regional Climate Services: Integrating Climate Knowledge and Ensemble Projections

Authors: Daisuke Matsuoka1*, Kōshirō Murakami1**, Ryō Matsumoto1, Rui Itō1, Shiori Sugimoto1, Daisuke Sugiyama1, Masayuki Hara2, Masaaki Hayashida3**, Nguyen Trung Kien3**, Aurélie Peng3, Daishi Abe3, Kazunari Sugiyama3

Affiliations: 1. Japan Agency for Marine-Earth Science and Technology, 2. Kochi University, 3. Ridge-i Inc. *Corresponding author, **At the time of research

DOI: https://doi.org/10.1029/2025JH001205

[Glossary]

※4 Benchmark

A common test used to objectively evaluate the performance of AI models. It quantifies a model's knowledge and reasoning abilities, serving as a guide for selecting the optimal model for a given purpose.

3. Background

With the progression of global warming, the frequency and intensity of extreme weather events such as heatwaves, heavy rainfall, droughts, and sea-level rise are increasing. To address these challenges, it is essential to scientifically assess future climate risks and rapidly plan and implement "adaptation measures" tailored to the specific conditions of each region. Local governments, at the forefront of climate change adaptation, play a central role in formulating region-specific action plans. However, developing effective adaptation plans requires not only climatological expertise but also interdisciplinary knowledge in areas such as regional industry, public policy, and economics, as well as advanced data analysis capabilities. This hurdle is particularly high for local governments with limited specialized personnel and financial resources, raising concerns about a widening gap in adaptation capacity between regions.

LLMs, which have been rapidly evolving in recent years, are expected to provide access to specialized knowledge through natural language. However, general-purpose LLMs, primarily trained on common web text, lack accurate specialized knowledge in climate science and carry the risk of generating plausible but incorrect information (hallucinations). Furthermore, the direct analysis and utilization of quantitative numerical data, such as "future climate projection data" essential for risk assessment, have been technically challenging for LLMs.

4. Achievements

A joint research team from JAMSTEC, Kochi University, and Ridge-i Inc. has developed a region-specific LLM capable of integrating and utilizing climate science expertise and quantitative future projection data. This research adopted the open-source LLM "Llama 3.3 Swallow 70B Instruct v0.4," which has strong Japanese language capabilities, developed by Tokyo University of Science, as the base model. This model underwent specialized fine-tuning (※5) for climatology using 338 academic papers on climate change adaptation registered on the Climate Change Adaptation Information Platform (A-PLAT) operated by the National Institute for Environmental Studies, and IPCC (Intergovernmental Panel on Climate Change) assessment reports.

Furthermore, we advanced Retrieval Augmented Generation (RAG) technology, which utilizes external knowledge, to build a system that can automatically search and extract quantitative numerical data from the "Ensemble Climate Projection Database for Promoting Global Warming Countermeasures (d4PDF)" along with textual data such as regional adaptation plan guidelines, based on user queries (Figure 1).

Performance evaluation of the developed model using a climatology-specific benchmark showed significant improvements in both Japanese and English compared to the base model, demonstrating particularly excellent capabilities in highly specialized fields such as "Impacts, Adaptation, and Vulnerability" and "Mitigation Measures" (Figure 2). Additionally, as a case study for PoC, we conducted a study on planning heat countermeasures in Kumagaya City, Saitama Prefecture, which faces challenges with extreme high temperatures. The system extracted temperature rise values from future climate projection data (RCP8.5 scenario) and dynamically searched rules from Kumagaya City's guidelines, such as "criteria for installing green curtains and rest areas based on the increase in heatstroke patients." It then successfully proposed quantitative requirements for infrastructure expansion for each case (average, optimistic, and pessimistic) based on probabilistic ensemble projection data, while transparently showing the calculation process. Furthermore, we demonstrated the system's ability to integrate discussions into actionable plans by having the LLM simulate different expert roles such as "scientist," "consultant," and "local government official," considering the balance of effectiveness, cost, and feasibility.

Figure 1: Processing flow in the system using the region-specific LLM

Users input instructions and questions in natural language to a chatbot-type application. If necessary, the system retrieves relevant contextual information, such as future projections and current adaptation measures, from databases containing quantitative climate projection data and past regional adaptation measures through semantic search. The system then combines the extracted information with the user's query to instruct the LLM (prompt) and presents the generated response, based on specialized knowledge and projection data, to the user.

Figure 2: Comparison of AI model accuracy in the field of climate change

In a test of specialized knowledge in climatology, the model developed in this research demonstrated high performance, exceeding the correct answer rates of general-purpose LLMs such as Swallow 70B and GPT-4o in almost all areas.

[Glossary]

※5 Fine-tuning

A technique where an already trained AI model undergoes additional learning with data specialized for a particular field or task.

5. Future Outlook

This research has demonstrated the technical foundation for enabling practitioners in local governments and SMEs, who may lack advanced expertise or abundant resources, to conduct data-driven scientific climate risk assessments and develop adaptation measures with AI assistance. Crucially, AI functions not as a complete replacement for human decision-making but as a preliminary support tool that rapidly presents diverse response scenarios from vast data, strongly aiding human deliberation and consensus-building.

This framework has global applicability beyond Japan. By replacing the reference database in RAG with climate data and socio-economic information of the target region, it can be customized to provide region-specific climate services to various areas, including developing countries vulnerable to climate change, without the need for costly retraining.

As a next step, we will collaborate with institutions promoting climate change adaptation in Japan, such as the National Institute for Environmental Studies and various local governments, to develop a service that allows anyone to perform expert-level analysis and countermeasure planning.

The widespread adoption of such next-generation regional climate services, powered by scientific data and AI, is expected to contribute to reducing economic losses from climate change and realizing a safe and resilient society.