Ryan Lagerquist-April 20

Deep Learning for Detection of Synoptic-Scale Fronts


Lagerquist, Ryan
Ph.D. Student


April 20, 2018 - 3:30 pm


April 20, 2018 - 4:00 pm


120 David L. Boren Blvd., Room 5600, Norman, OK 73072   View map

Deep Learning for Detection of Synoptic-Scale Fronts

Deep learning is a subset of machine learning, in which the model builds represen- tations of input data at various levels of abstraction. Deep learning has exploded in popularity in the last few years, thanks to advances in computing resources and opti- mization algorithms. Convolutional neural networks (CNN), the most common type of deep-learning model, are specially designed to learn from data with topological structure, such as spatiotemporal grids. I used CNN’s to draw synoptic-scale warm and cold fronts in the North American Regional Reanalysis (NARR). The predictor variables (input grids) are θw and horizontal wind. The verification data consist of human-drawn fronts from the Weather Prediction Center (WPC) surface bulletins.


Models are judged via neighbourhood evaluation, based on predicted warm-front and cold-front probabilities at each grid cell, and object-based evaluation, where predictions are explicitly turned into polylines (analogous to human-drawn fronts).

For neighbourhood evaluation – where grid cell (i, j) is compared to all grid cells in a 3-by-3 neighbourhood around (i, j) – the best model achieves a Peirce score of 0.704,

Gerrity score of 0.700, and probability of detection of 0.896, all significantly better than climatology. However, it predicts fronts ~16 times more often than they occur, which is an unacceptable bias. For 100-km object-based evaluation, this frequency bias is reduced to 1.11 (slight overprediction), but the probability of detection is reduced to 0.220 (unacceptably low).


Despite the aforementioned problems (trade-off between probability of detection and bias), the model predictions are subjectively realistic and largely match human- drawn fronts, which suggests that with further improvements this tool could be used operationally. In future work I will explore other deep-learning and object-conversion methods, with the goal of finding a better-trade off between probability of detection and bias. Also, I will apply deep learning to the spatial prediction of tornadoes and damaging convective wind on the warning time scale (0-2 hours).