AI-Powered Climate Action Prioritization

Cities are centers of innovation, culture, and economic activity playing a crucial role in driving large-scale environmental change, as they account for over 70% of global greenhouse gas emissions. While many cities are aware of the need to reduce emissions, many still lack comprehensive, data-driven tools to prioritize impactful actions. This often results in uncertainty and delays, slowing implementation of effective climate measures.

Our software, CityCatalyst, helps cities address climate change by building greenhouse gas (GHG) inventories and prioritizing effective climate actions. It simplifies complex climate data, integrates emissions and risk insights, and uses AI tools to guide cities—especially those with limited resources—toward impactful, data-driven decision-making.

In collaboration with our partners, I Care Brasil and Brisa Soluções, we launched  a project with  50 cities across Brazil, home to 25 million people. This initiative is part of the Green Resilient Cities Program, supported by the Bloomberg Philanthropies Joint Program (BPJP), the Global Network of Mayors (C40), the Global Covenant of Mayors (GCoM) and CDP. Our goal was to develop a software tool to help these cities select the most impactful and appropriate climate actions tailored to their unique circumstances. Recognizing the diverse emissions profiles, climate risks, needs, and challenges across cities, we created the High Impact Action Prioritization (HIAP) tool. This tool lowers entry barriers, enabling cities to begin their climate action journeys with a tailored approach that fits their specific needs.

This project embodies the  mission of AI for Good—harnessing  artificial intelligence not just for innovation, but for meaningful, equitable, and scalable climate solutions. Our work showcases  how AI, when responsibly developed and applied, can empower local governments to act effectively in the face of climate change, demonstrating that AI can be a force for good in addressing global climate challenges.

A Holistic Approach to Prioritization:

Our HIAP tool integrates city-specific emissions data, climate change risk assessments (CCRAs—see our related blog post), and city-specific context to create a tailored list of climate actions. Here's our approach:

Step 1: Consolidating Climate Action Lists:

The first step in creating individualized climate action recommendations is understanding available actions, their impacts (e.g., sectors affected), budget considerations, implementation timelines, additional co-benefits (or adverse side-effects) and more. We begin by compiling a comprehensive long list of climate actions from established frameworks and databases, including:

• C40’s Climate Action Library
• Intergovernmental Panel on Climate Change (IPCC) reports
• Consolidated expertise from our local partner I Care

Actions are categorized by sector (e.g., stationary energy or transportation) according to the Greenhouse Gas Protocol for Cities (GPC), hazard (e.g., floods or heat waves), and type (mitigation or adaptation). A standardized data schema ensures consistency across the three diverse sources, capturing key details such as GHG reduction potential and adaptation effectiveness. While this database serves as a foundational resource, it is not exhaustive.

Step 2: Expert Input for AI Model Training:

While having a list of potential actions is important , determining the most impactful action for a specific city is complex due to varying contexts and long-term effectiveness. Actions implemented today may take  years to fully make an impact, which makes assessment even more challenging. To better understand these complexities, we conducted workshops with climate experts from the C40 network, using our Climate Action Labeling Tool to gather their insights.

The tool employs a pairwise comparison method, where experts evaluate climate actions in the context of specific cities to identify the most impactful choices. This expert-driven data serves as direct input for training our AI models, enhancing their decision-making capabilities.

‍

‍

Explore our labeling tool here: Climate Action Labeler*

_{*This tool was used in the initial stage of the project for a selected group of cities. The data has not been updated and may not reflect the current state of those citie’s inventories and climate risks. Additionally you might experience downtime with this tool while being worked on or under maintenance.}

‍

Step 3: AI-Enhanced Action Scoring:

With our climate action database and expert-derived scoring approach established, we built our AI tools to help rank  actions for cities. Initially, we explored various heuristic models, large language models (LLMs), and classical machine learning (ML) approaches like XGBoost. Each of these has  its own advantages and drawbacks. In general, heuristic models are easy to understand, implement, explain and they can work with limited or even no training data. However, they can be overly simplistic and struggle with complex problems. LLMs excel at handling  unstructured data and  capturing complex relationships and context, but they are difficult to interpret and can struggle to pinpoint decision making. Additionally, they require large amounts of data for domain specific finetuning. 

Classical ML models, on the other hand, are highly effective with tabular data, offer good interpretability and perform well with less training data. However, they require complex data preprocessing and have very limited capabilities for processing unstructured data.

‍

‍

Given the limited amount of training data and its primarily tabular nature, we found that XGBoost delivers the best performance. Due to the small training dataset, we had to significantly reduce the high-dimensional feature space by compressing dozens of individual features into a smaller set of aggregated ones. The model uses engineered features from both the climate action pair and the city, and outputs the preferred action based on the learned feature importance from the training data.

Before making a prediction, we apply some prefiltering to the actions based on specific attributes, such as the citie’s biome. This ensures the suggested actions are relevant and aligned with the city’s characteristics. The model is a classification model that outputs one of two possible classes: 0 or 1. A class 1 indicates  that the first action is preferred, while a class 0 indicates that the second action is preferred. For each feature, the model computes a score, and the overall score is determined by summing up all the individual scores. The magnitude of each score is influenced by the difference in feature values and the learned weights of the model. The following graph illustrates the model’s decision-making process. In this case, the final score is 0.573, shown in the top right corner, which means the predicted class is 1.

‍

‍

Our trained AI model effectively scores climate actions based on city-specific contexts, identifying the most suitable actions for further review and selection. A key challenge however, is to effectively integrate unstructured data into the models. To address this, we are exploring  hybrid approaches involving LLMs, which can leverage unstructured data—common in complex city contexts that aren't always easily quantifiable or categorizable.

Step 4: Generating an Implementation Plan:

Once actions are ranked, cities can generate detailed implementation plans for selected climate actions using our automated plan creator tool. To facilitate this, we established a vector database containing nearly 100 documents, ranging from city climate action plans and case studies of implemented actions to national climate strategies of Brazil.

We implemented an agentic workflow to craft step-by-step actionable plans, using contextual information from the database. This involves  extracting the most relevant documents for a given city and action, which then  informs the agentic workflow. Each implementation plan includes:

• Description and alignment with national strategy
• Sub-actions for implementation
• Involved municipalities
• Milestones
• Measurement, Evaluation, and Reporting (MER)
• Addressed climate hazards
• Addressed mitigation sectors
• Alignment with SDGs

We plan to expand these sections further, focusing on critical areas like financing to enhance a cities' ability to fund and quickly implement actions.

Planned Improvements to our Tool:

Our current version demonstrates the potential of combining data and AI to drive positive scalable impact in cities. We aim to enhance climate action by improving data quality through better alignment across various sources and refining existing descriptive fields, such as emission reduction potential, cost, timelines, and technical feasibility. Additionally, we are committed to continuously expanding our action database and introducing more comprehensive features, including financing tools.

Key Insights and Your Feedback for Improvement:

The HIAP tool represents a significant advancement in climate action planning, combining  sophisticated AI with actionable, local insights for city sustainability and resilience. However, challenges like obtaining accurate and comprehensive data persist, particularly when it comes to acquiring ‘ground truth’ data. In this context, ‘ground truth’ refers to data that is verified to be objectively and factually correct, often through direct measurement or authoritative sources. 

The impacts of many climate actions are uncertain, assessable only in the long term, and highly specific to each city's unique circumstances, making it difficult to generalize findings to other cities. While expert feedback is valuable, it may still be influenced by personal judgment or incomplete information, meaning it doesn't always reflect the same level of certainty. 

We have observed differing  preferences among experts when comparing identical city-action pairs. The lack of factual ‘ground truth’ data introduces challenges and uncertainty into the automated ranking process. 

If you have experience with similar projects, climate actions, or suggestions for improvements, we would love to hear from you! Feel free to contact us at info@openearth.org. 

‍

Curious? Explore our tool here: Climate Action Prioritization*

^{*The tool is under active development and may experience downtime, feature changes, or data changes at any moment.}

‍