How AI can help predict climate extremes

The threats of climate change are real.

Record-shattering temperatures in the US Pacific Northwest this month would have been all but impossible without climate change. It's the same for the 2020 Siberian heat wave, which accelerated the thawing of carbon-rich permafrost to an alarming rate. Catastrophic wildfires, extreme rainfall and flooding are becoming more frequent and intense across the globe, with each extreme event increasingly exhibiting the fingerprints of anthropogenic climate change.

As part of IBM's mission to accelerate solutions for climate change, we are driving new research to build real and impact-focused solutions for organizations to model and assess their climate risk and mitigation strategies. This week, our group showcases nine new research papers at the International Conference for Machine Learning (ICML) 2021: Tackling Climate Change with Machine Learning—including one selected for a spotlight, an honor given to less than 10% of accepted submissions.

Here we highlight three of our contributions.

Quantification of carbon sequestration in urban forests

Trees can extract CO₂ from the atmosphere and store it in their trunks, stems, and roots. But not all trees are equal, as some will store carbon faster but may have a short lifetime; while other trees may extract CO₂ slower but live much longer. It is important to quantify carbon sequestration in trees across cities or forests, as climate change can impact where and how fast trees grow in a geographical area and how much carbon can be stored in them locally.

To address this unmet need, our ICML paper¹ used hyperspectral information from the PAIRS Geospatial combined with image segmentation and deep learning classification tools to develop a carbon calculator that quantifies carbon storage at the individual tree level. Combining climate, weather, satellite, and Lidar data on the PAIRS platform, the optimum location for tree planting can be estimated across large regions and a tree’s capacity to extract and store carbon can be quantified multiple times per year.

Our approach supports the needs of emerging carbon trading markets: transparent tools and verifiable information to quantify carbon sequestration and offset greenhouse gas emissions. Using open source imagery in combination with deep learning tools, a transparent method is proposed that may be employed to determine optimal tree species planting patterns, maximize the carbon sequestration, and enable businesses to offset their carbon emission and achieve net zero carbon emission.

Predicting extreme rainfall with transformers

Extreme rainfall can be disruptive, damaging and devastating. But accurately predicting its likelihood beyond a few weeks remains a significant challenge, even for state-of-the-art seasonal forecasting systems. In our ICML spotlight-selected paper,² we applied a transformer-based machine learning architecture to directly predict maximum daily precipitation in a given week—up to six months into the future.

The temporal fusion transformer (TFT) network combines multi-horizon forecasting with specialized components to select relevant inputs and suppress unnecessary features. It is unique in that it can also seamlessly integrate predictions from external sources – a challenging task in time series predictions. We trained the TFT network with both high-fidelity historical data from the PAIRS platform and physics-based forecasts from IBM's Seasonal Probabilistic Forecasting Platform to create a unique physics-guided AI prediction model for extreme rainfall.

Our results show that our TFT network consistently outperforms existing state-of-the-art forecast systems, even at six month time horizons. While this appears to be ready for celebration, another key benchmark of a successful seasonal forecasts out to six months is whether it can beat climatology. Here, climatology means the average conditions over some prior period, usually 10 to 30 years.

Seasonal forecast models rarely improve upon the skill of climatology beyond four to five months into the future. In other words, we may as well use the average of historical weather conditions to predict what will happen this November. However, we are thrilled to report that the TFT network also outperforms climatology in key locations around Florida and Rio de Janeiro—our two test areas—promising an exciting future for IBM's AI-powered seasonal extreme rainfall predictions.

Choose your own (weather) adventure

Climate change is making extreme weather more extreme. Given the inherent uncertainty of long-term climate predictions, many sectors – including finance, energy, asset management, logistics and agriculture—also use plausible climate “what-if” climate scenarios to explore their own risk exposure, resilience and mitigation strategies.

Since the 1980s, these “what-if” climate scenarios have been created using stochastic weather generators. However, it is very difficult for traditional weather generation algorithms to create realistic extreme climate scenarios because the weather data being modeled is highly imbalanced, contains spatiotemporal dependencies and contains extreme weather events exacerbated by climate change.

Our ICML paper³ shows that variational auto-encoders (VAEs) offer a pathway for efficient and controllable climate scenario synthesis, especially for extreme weather events. VAEs are an encoder-decoder generative model which we used for unsupervised clustering of weather events in the latent space and subsequent generation of high-quality climate scenarios (see Fig. 3).

The novel, unsupervised clustering of weather events and subsequent generation of high-quality climate scenarios represents a potentially important step forward in AI and climate science. We expect it to provide unparalleled flexibility in high-fidelity weather generation, especially for extreme climate events.

And in a new partnership with the University of Illinois Urbana-Champaign, we will apply computational creativity algorithms we have used in the past with recipe developers to extrapolate our deep generative weather generators to unseen future climate scenarios.