By Dave DeFusco

With the help of AI, scientists can better model climate patterns, identify trends and make predictions, leading to a clearer understanding of climate change and its environmental impact. Katz School students in Artificial Intelligence and Data Analytics did their part for S&P Global Market Intelligence by sifting through large amounts of data, like satellite images that researchers use to monitor climate change, to create a set of geography-specific, super-resolution climate forecasts that will be valuable for climate research, policy planning and risk assessment.

“This project represents a significant step in the application of deep learning and AI techniques to climate modeling,” said Dr. Yuri Katz, an industry professor of mathematics and a group manager of data science at S&P Global Market Intelligence. “It generated high-resolution climate forecasts, which can provide more detailed insights into the potential local impacts of climate change under different scenarios.”

Under the supervision of Dr. Katz and Data Analytics Industry Professor James Topor, the students worked on an S&P Global project that leveraged temporal and spatial climatological data, specifically global precipitation and humidity measurements, from two satellites of NASA’s Earth Observing System.

Working alongside S&P Global data scientists, the student broke into two teams—one focused on precipitation and the other on humidity—and collected data, conducted exploratory data analysis and helped develop and test models. The precipitation team included Haider Ali, Marwan Kanaan, Jatin Kayasth, Kanchan Maurya and JuJu Ren, and the humidity team included Varan Bhatnagar, Aishwarya Deshmane, Deepa Paikar, Reiyo Reiyo, Avinash Swaminathan and Manish Thota.

“Precipitation and humidity required two different data collection methods, databases and resolution of measurements,” said JuJu Ren, a Katz School master’s candidate in AI who led the precipitation team. “Precipitation may lead to pluvial flooding, whereas growing humidity, combined with elevated temperature, may lead to a higher heat index.”

Dr. Katz said the goal was to create high-resolution versions of General Circulation Model (GCM) climate forecasts. GCM is a computer-based mathematical model used in the field of atmospheric and oceanic sciences to simulate and study the behavior of the atmosphere, oceans and climate systems based on Shared Socioeconomic Pathways (SSP) CMIP6 scenarios.

SSP is a set of scenarios that climate researchers use to explore possible global futures based on varying socioeconomic and environmental factors, and CMIP6, or the Coupled Model Intercomparison Project Phase 6, plays a crucial role in supporting the Intergovernmental Panel on Climate Change assessment reports, which are instrumental in shaping global climate policy.

To complement the satellite data, the teams collected data from the NASA Earth Exchange Global Daily Downscaled Projections (NEX-GDDP), which offers high-resolution climate projections based on different scenarios—crucial for climate modeling. An exploratory data analysis was performed to assess the completeness and alignment of the satellite and NEX data, which was vital for ensuring the quality and compatibility of the data.

The satellite data was preprocessed using various techniques such as aggregation and rescaling. Deshmane, leader of the humidity team, said her team enhanced the resolution of climate data, specifically NASA’s humidity estimates, by a factor of five. To ensure the reliability of their model, they conducted a thorough validation process by comparing their model’s outcomes with data from previous years.

“To accomplish this, we harnessed the power of neural-network, computer-vision techniques, which allowed us to refine the spatial resolution of the NEX climatic projections,” she said. “This enhanced level of detail will prove instrumental for S&P’s research team in making more precise predictions regarding changes in humidity.”

The teams then developed a set of geography-specific, super-resolution deep learning models, which are designed to enhance the resolution of climate forecasts for specific regions available in the NEX GDDP product. “The outcome of the project was the development of a machine-learning data pipeline, which supported the entire process, from data preprocessing and model training to validation, testing and model deployment,” said Dr. Katz.