Forecasting the environmental change

The Arctic is warming four times as fast as the rest of the planet. Even if we can limit our emissions to less than 2 degrees Celsius below pre-industrial levels, Arctic warming could increase by 4-5 degrees Celsius due to complex positive feedback mechanisms. This could have serious consequences for the region and the planet. Perhaps the biggest change to the region over recent decades has been the loss of almost half the summer sea ice since the early 1980s from the Arctic Ocean. The most substantial change however is the loss of the thick sea ice that usually circulates within the Arctic Ocean for many years, known as multiyear ice. The sensitivity of sea ice to increasing temperatures has caused the summer Arctic sea ice area to see massive shrinkage over the past forty years, equivalent to the loss of an area around 25 times the size of Great Britain. Climate projections show that by the middle of the century, the Arctic could have seen its first ice-free summer. These accelerating changes have dramatic consequences for our climate, Earth’s ecosystems, and the lives of Indigenous and local communities whose livelihoods are tied to the seasonal sea ice cycle.

Reliable sea ice forecasting systems have the potential to help conservation workers, local communities, and research expedition planners anticipate changes in sea ice conditions. Researchers, meteorologists and climatologists depend on rich historical data from different sources to make high-resolution forecasts for a certain region. Modern forecasting techniques use mathematical models of multiple atmospheric conditions that can accurately project longer-term climate predictions (reference: Brenner, Laurie. Four Types of Forecasting). Even with large datasets, physics-based predictive models and advanced computers to run them on, it is, nonetheless, extremely difficult to forecast conditions in remote and extreme environments due to their complex relationship with their surroundings in a rapidly changing climate. AI has the potential to transform forecasting techniques by enhancing the predictive ability of forecasts for a wide range of environmental phenomena, serving as an early warning system for local populations living close to sea ice.

Combining large-scale in situ and remote sensing data from environmental monitoring systems at different scales collected over a long period, AI algorithms can be trained to accurately forecast long-term and short-term fluctuations in vulnerable regions. Additional data from physical and digital infrastructure, as well as various open data, citizen science and crowdsourcing efforts can further enhance the accuracy of forecasting tools. Applications of these AI tools could range from providing accurate daily weather forecasts, forecasts of a city’s energy consumption, the wider development of crisis management plans, data-assisted policymaking, timely interventions for supporting communities in high-risk zones, local capacity for renewable energy and protection for native fauna or wildlife.

AI Method in focus: Deep Learning

The average levels of Arctic sea ice in summer months have halved since satellites began monitoring sea ice in 1979. Interactions of the sea ice environments are challenging to capture using purely physics-based models, which led researchers to explore advanced deep learning-based AI solutions to forecast sea ice changes.

Deep learning is a subfield of ML that uses a complex system of mathematical algorithms called an artificial neural network, inspired by the biological neural network of the human brain. Data sources such as satellites, ground-based sensors and other similar devices routinely generate a large amount of data that are too vast and complicated to draw meaningful conclusions from without sophisticated computational solutions. Deep learning uses a layered approach in its computations, giving AI tools the ability to process unstructured data from multiple data sources to conclude with little to no human intervention. Although deep learning requires large datasets and training models can take days or weeks, once trained, deep learning systems run significantly faster than their physics-based counterparts and reduce the testing time by multiple orders of magnitude.

We recommend Deep Learning References by Pablo Mesejo from Inria Grenoble Rhone-Alpes, Perception team from 2017, and a more recent and up-to-date references on GitHub repository Awesome Deep Learning papers by Terry T. Um to gain technical details on Deep Learning.