Advanced Machine Learning and Data Mining

Physics-informed two-stage framework for sub-seasonal tropical cyclone genesis (TCG) prediction. (A) A TimeSformer classifier analyzes 16-day sequences of nine physics-based predictors and estimates genesis probability on overlapping 20°×20° patches. (B) A U-Net then refines only the flagged patches into pixel-level probability maps of likely genesis locations. The design encodes the two-phase physics—thermodynamic preconditioning followed by mesoscale organization—improving temporal and spatial precision while reducing false alarms and computational cost.

Anthropogenic global warming has triggered potentially dangerous shifts in climate and ocean systems, leading to more severe and frequent climate extremes.

To minimize the potential damage from these extremes in a warming climate, we should enhance our predictive capabilities through the development of innovative data-mining techniques.

In our research, we pioneer new data-mining methods, including supervised dimensionality reduction coupled with machine learning techniques, to improve the predictability of hydro-climate extremes in the context of statistical downscaling and predictions.

Related Papers

Ali Sarhadi, Donald H. Burn, Ge Yang, Ali Ghodsi, (2017), “Advances in projection of climate change impacts using supervised nonlinear dimensionality reduction techniques”, Climate Dynamics, DOI 10.1007/s00382-016-3145-0. Read Article

Ali Sarhadi, Donald H. Burn, Fiona Johnson, Raj Mehrotra, Ashish Sharma, (2016), “Water resources climate change projections using supervised nonlinear and multivariate soft computing techniques”, Journal of Hydrology, Vol: 536, 119-132. Read Article