J61.4 Integrated Climate Extremes: Modeling Future Impacts for Visualizing Climate Change

Thursday, 16 January 2020: 9:15 AM
Surya Karthik Mukkavilli, Montreal Institute for Learning Algorithms (Mila), Montreal, QC, Canada; and Y. Min, A. Madanchi, V. B. Pacela, S. Patel, and Y. Bengio

The public awareness of, and concern about, climate change often does not match the magnitude of its threat to humans and our environment. One reason for this mismatch is that it is difficult for people to mentally simulate the effects of climate change, which is an inherently complex process with probabilistic outcomes. To overcome these challenges, the public needs an easily accessible tool to help them visualize climate change [1], both rationally and viscerally, by bringing the future closer with predictions of their local environment (for example, showing an individual the frequency and impact of events like floods at an address of their choice). This visualization tool to depict accurate and personalized predictions of climate change requires combining cutting-edge techniques from artificial intelligence, psychology, gaming simulations, economics and climate science. This work will focus on the integrated climate extremes work within the visualization tool, used to model future extreme climate impacts such as floods. The results over Canada and the USA, from the integrated flood hazard prediction system developed for the year 2050/2100 will be demonstrated. The future probabilistic precipitation predictions are based on a stochastic prediction system with historical precipitation and forcing data from atmospheric constituents, a future scaling model for forced climate variability, Bayesian statistics to estimate free parameters, maximum likelihood approaches to represent internal climate variability with fractional Gaussian noise [2, 3]. Return period of precipitation fluctuation for each region [4], and their future return periods due to increase in mean value of projected precipitation [5] are produced. A future snow and glacier melt prediction model from multi-spectral satellite images with stochastic video generation techniques using an LSTM and deep convolutional GAN [6] has been developed. Streamflow predictions are made with spatio-temporal LSTM models. The integrated flood hazard system then regrids and combines the snow and precipitation predictions, along with runoff information, to predict inland water levels with probabilistic and random walk based flow transitions for different elevations. Finally, this integrated flood hazard map is used to generate future visualizations of climate change for different houses in North American with GANs.

[1] Schmidt, V., Luccioni, A., Mukkavilli, S. K., Balasooriya, N., Sankaran, K., Chayes, J., & Bengio, Y. (2019). Visualizing the consequences of climate change using cycle-consistent adversarial networks. arXiv preprint arXiv:1905.03709.

[2] Amador, L. D. R., & Lovejoy, S. (2019). Predicting the global temperature with the Stochastic Seasonal to Interannual Prediction System (StocSIPS). Climate Dynamics, 1-39.

[3] Hébert, R., & Lovejoy, S. (2018). Regional Climate Sensitivity‐and Historical‐Based Projections to 2100. Geophysical Research Letters, 45(9), 4248-4254.

[4] de Lima, M. I. P., & Lovejoy, S. (2015). Macroweather precipitation variability up to global and centennial scales. Water Resources Research, 51(12), 9490-9513.

[5] Kodra, E., & Ganguly, A. R. (2014). Asymmetry of projected increases in extreme temperature distributions. Scientific reports, 4, 5884.

[6] Denton, E., & Fergus, R. (2018). Stochastic video generation with a learned prior. arXiv preprint arXiv:1802.07687.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner