Artificial Intelligence for Climate Action in Developing Countries: Opportunities, Challenges and Risks

Isabelle Tingzon�Geospatial Data Scientist |The World Bank

Climate Change AI

​

• Climate change and digital transformation are the two most powerful trends of this century.

Climate Change & Digital Transformation

Artificial intelligence (AI): Any algorithm that allows a machine to perform a complex task (e.g. tasks associated with human intelligence)

Machine learning (ML): Techniques that automatically extract patterns from large amounts of data , which can then be used to make predictions/recommendations on new data.

• The rise of AI shows promise in supporting climate action, but may also carry risks itself and needs to be developed responsibly.

AI technologies can enable strategies to accelerate climate action when applied responsibly in partnership with relevant stakeholders.

Climate Change AI Report�Tackling Climate Change with Machine Learning (Rolnick et al., 2019)

Buildings and Cities

AI can help conserve building energy by enabling low-carbon urban planning, energy use modeling, and building function optimization, e.g. heating, lighting.

Source: Clutton-Brock, Peter, et al. Climate change and AI. recommendations for government action. GPAI, Climate Change AI, Centre for AI & Climate, 2021.

Case Study

AI for Energy Modeling and Conservation

Using ML to predict building use, height, and year of construction to estimate energy demand and enable better building energy conservation.

​

Sources: Milojevic-Dupont, Nikola, et al., 2023,; Wurm, Michael, et al., 2021.

Farms and Forests

AI can facilitate nature-based solutions, precision agriculture, estimate carbon stock, detect illegal deforestation, and accelerate afforestation.

Source: Clutton-Brock, Peter, et al. Climate change and AI. recommendations for government action. GPAI, Climate Change AI, Centre for AI & Climate, 2021.

Case Study

AI for Precision Agriculture

Researchers at NASA Harvest are using ML and time-series satellite images for crop type classification and yield estimation.

Source: https://www.catherinenakalembe.com/research/projects

Societal Adaptation

AI can help improve societal adaptation, for example:

• Improve public health models for climate-influenced diseases
• Disaster risk reduction: Identify vulnerable/at-risk population & infrastructure
• Disaster response: Rapidly detect damaged/destroyed structures

The Caribbean is among the most climate-vulnerable regions in the world.

​

​

In 2017, Hurricane Maria destroyed ~90% of Dominica’s housing stock with damages > 380M USD in the housing sector.

*Dominica Climate Resilience and Recovery Plan 2020-2030

Case Study

Climate resilience programs by national government agencies

​

​

• Disaster Vulnerability Reduction Project by the Government of Saint Lucia
• Resilient Housing Scheme by the Government of Dominica

Climate resilience initiatives require comprehensive housing stock data.

​

​

However, this data is often limited, incomplete, inaccessible or completely non-existent.

AI and Earth Observation can help fill in critical exposure data gaps.

​

​

The Digital Earth for Resilient Caribbean aims to enhance local capacity to leverage AI & Earth Observation for resilient housing operations.

Leveraging Earth Observation Datasets

Open Data

Restricted

VHR Aerial Images

Drone images, RGB orthophotos

LiDAR Data

DSM, DTM, nDSM

​

Building Footprints

delineated from the aerial images

​

Street View Images

taken using GoPro cameras mounted on cars

​

Summary Workflow for Roof Classification

Segment Anything Model

Convolutional Neural Network

Aerial Image

Building Footprints

Rooftop Image Tiles

Roof Classification Map

15

Aerial Imagery

LiDAR

AI-generated Map

Flat

Gable

Hip

Roof is pitched or sloped on 3 or more sides

Roof is pitched on two side up to a central ridge.

Roof is flat with a slope less than 7 degrees.

Roof Type Classification

16

AI-generated Map

Aerial Imagery

Concrete Cement

Roofs are made of concrete/cement

Healthy Metal

Includes corrugated metal, galvanized sheeting, and other metal material

Irregular Metal

Includes metal roofing with rusting, patching, or some damage. These roof carry a higher risk.

Incomplete

Under construction, severely damaged or haphazard

Blue Tarpaulin

Roof is covered with blue tarpaulins, indicating damage

Roof Material Classification

17

Exposure Data Layers

Roof Classification Maps

Generating roof type and roof material classification maps with 87-92% accuracies.

Saint Lucia

Grenada

Dominica

Pre- and Post-disaster Classification Maps

Open Data

Restricted

19

Street View Data Collection

GoPro cameras are mounted on top of cars and driven around the neighborhood to collect street view images.

Summary Workflow for Roof Classification

Detectron2

Convolutional Neural Network

Street View Photo

Building Outlines

Building Image

Building Classification

Residential, Complete, Plaster, One-story building

Residential, Incomplete, Brick or Cement, Two-story building

Attributing Street-Level Characteristics to Building Footprints

Residential, Incomplete,

Brick or cement, One story

Residential, Complete,

Plaster, Two story

Residential, Complete,

Plaster, Two story

AI as a tool for climate action

Data and Digital Infrastructure

Challenges

• Data Scarcity. Data required for AI applications (e.g. field surveys) are often limited, incomplete, inaccessible, or completely non-existent.
• Unequal data availability. Data collection is often concentrated in the Global North, potentially leading to biased models/systems.
• Privacy, security, & reputational risks. Incentives for organizations to share data are often outweighed by the costs and risks of doing so.

​

Recommendations

• Fundamental data governance: Support governments in investing in the digital infrastructure needed for the consistent collection, management, and processing of large volumes of data.
• Eliminate data silos and support the development of standards and protocols for data sharing.
• Create initiatives to increase data sharing and access. Data portals for easy access to climate-relevant datasets such as the Risk Data Library.

AI as a tool for climate action

Data and Digital Infrastructure

Challenges

• Biases can arise from disparities in data collection
• Model calibration may be optimized for particular regions
• Need to continually interrogate AI systems and understand their implications for decision-making.

Mitigating biases in data and models

Biases in data and models

Representation Bias

Occurs when the data used to train the model isn’t representative for the problem that needs to be solved.

​

​

Many large-scale datasets are distributed across more the Global North and often generalize poorly to developing countries.

Example: The most widely used image-recognition systems are better at identifying items from wealthy households than from poor ones.

de Vries, Terrance, et al. "Does Object Recognition Work for Everyone?." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. 2019.

• Need for localized solutions → inclusive solutions that incorporate local data, context, and perspectives.
• Bridge the gap between AI experts and climate-relevant sectors. Need an interdisciplinary approach that draws insights from domain experts, policymakers, and affected communities.
• Strengthen local capacity. Implement AI capacity building programs for policymakers, industry leaders, and civil society.
• Better understand requirements, capabilities, limitations, and risks of AI solutions

Local Capacity Building

Recommendations

Responsible AI in the context of climate action

Avoid techno-solutionism

• AI is not a silver bullet solution
• AI solutions should be informed by societal contexts
• AI is only a component of the solution and not a solution in and of itself.

Mitigating biases and risks in AI

• E.g., Buildings data: Housing discrimination, uneven availability across geographies
• E.g., Weather models: Calibration may be optimized for particular regions

AI can have both positive and negative impacts on the environment.

• Energy use from compute-intensive resources → negative impacts
• Quantify negative & positive impacts of AI development
• e.g. Code Carbon for tracking/reducing emissions from AI model development

​

Thank you!

tisabelle@worldbank.org