"Utilizing Interpretable Neural Networks for Subseasonal Prediction"

Abstract: Within the field of atmospheric science, there is a common notion that neural networks are ‘black boxes’ or uninterpretable. However, recent advances in neural network interpretability allow for scientists to see into the ‘black box’ and thus, gain scientific insight into non-linear relationships. Here, we demonstrate the ability of a simple artificial neural network to predict midlatitude circulation on subseasonal timescales (2-5 weeks) using tropical storminess over the Indian and western Pacific Oceans. In addition, we utilize the interpretability technique known as layer-wise relevance propagation (LRP) to uncover favorable tropical conditions for subseasonal prediction.

Bio: Kirsten Mayer is a PhD student at Colorado State University (CSU) advised by Elizabeth Barnes. For her PhD work, she applies machine learning techniques to find and better understand favorable tropical conditions for subseasonal (2-5 weeks) prediction of the midlatitudes. Her broader interests include tropical-extratropical teleconnections, subseasonal to seasonal prediction as well as machine learning and its interpretability. Kirsten received a B.S. in Atmospheric and Oceanic Sciences from the University of Wisconsin-Madison and an M.S. in Atmospheric Sciences from CSU.

"Causal inference and causal discovery to study teleconnection pathways"

Abstract: Teleconnections refer to recurrent large-scale pressure patterns with low-frequency variability, connecting far-away geographical regions. They reflect modifications of atmospheric circulation affecting e.g. the position of the jet stream, storm track intensity or Monsoon strength and thus have a strong impact on our weather. However, extracting the physically relevant teleconnection pathways from observation or model data remains challenging. One major issue is to separate the signal from the noise given large internal atmospheric variability. This is compounded by varying dimensions in space and time and competing effects of different processes. Here, we discuss how novel data-driven causal methods beyond the commonly adopted correlation techniques can help to overcome some of these current limitations. We give an overview of causal inference frameworks and identify promising application cases common in climate science.

Bio: Marlene Kretschmer is a Postdoctoral Research Assistant (ACRCC) in the Department of Meteorology at the University of Reading. She earned a PhD in Climate Physics from the Potsdam Institute for Climate Impact Research & Potsdam University in 2018 , dissertation title: "Disentangling Causal Pathways of the Stratospheric Polar Vortex - A Machine Learning Approach" and an MSc in Mathematics at Humboldt University Berlin in 2014.

"Explaining Deep Learning Classification of Future Convective Storms"

Abstract: Deep convolutional neural networks (CNNs) can skillfully perform classification tasks because they can capture nonlinear and spatially invariant relationships among input features. CNNs have already proven skillful in atmospheric and climate science applications, including predictions of phenomena of varying temporal scales, such as severe thunderstorms and climate oscillations. However, as the climate continues to change, it is possible that extreme atmospheric and climate events will not be captured by deep learning models since the environmental conditions can change substantially from those in the past that were used for training the models. Here we evaluate the ability of a deep CNN to generalize and capture outlier convective storms of a future and warmer climate.

A deep CNN was trained to classify strongly-rotating convective storms extracted from a high-resolution climate simulation of a current climate period (2000-2013) over the continental United States. The performance of the model in classifying strongly-rotating storms extracted from a similar future climate simulation (RCP8.5) was then evaluated using performance diagrams and the Brier skill score. Results show that the model is able to classify storms with similar skill in both climates. The model is also skillful when evaluating a subset of future climate storms that contain very high low-level moisture content that exceeds that of current climate storms. Permutation feature importance reveals that zonal and meridional winds at 3-km above ground level are important variables for the classification task, which partly explains the model skill in classifying organized storms of a future climate. While not ranked as highly, thermodynamic variables also exhibit some importance, suggesting that the CNN is potentially learning about the importance of moisture and vertical instability in organized convection. Spatially, saliency maps show that gradients activate across focused areas near mesocyclones for strongly-rotating storms, while patterns are less organized for weaker convection. These results show that deep learning can generalize in a warmer and more moist climate, but questions are raised regarding the robustness of explainability methods used herein and whether these results would be consistent in a future climate simulation with an unconstrained synoptic-scale pattern.

Bio: Maria J. Molina is an Advanced Study Program Postdoctoral Fellow at the National Center for Atmospheric Research in Boulder, Colorado, hosted by the Computational and Information Systems Lab and the Mesoscale and Microscale Meteorology Lab. Maria’s research is at the intersection of severe convective storms, climate change, and deep learning. She earned a Ph.D. from Central Michigan University, a master's degree from Columbia University, and a bachelor's degree from Florida State University. In addition to atmospheric and climate research, she is passionate about effective science communication, diversity, equity, and inclusion, and early career scientist issues.

"Towards Physically-Consistent, Data-Driven Models of Convection"

Abstract: Data-driven algorithms, in particular neural networks, can emulate the effect of sub-grid scale processes in coarse-resolution climate models if trained on high-resolution climate simulations. However, they may violate key physical constraints and lack the ability to generalize outside of their training set. Here, we show that physical constraints can be enforced in neural networks, either approximately by adapting the loss function or to within machine precision by adapting the architecture. As these physical constraints are insufficient to guarantee generalizability, we additionally propose to physically rescale the training and validation data to improve the ability of neural networks to generalize to unseen climates.

Bio: Tom Beucler is an assistant project scientist working at the intersection of atmospheric physics, machine learning, and environmental fluid dynamics. He is affiliated with the University of California, Irvine, and Columbia University. He obtained his Ph.D. in atmospheric science from MIT, for which he was given the Rossby award for best doctoral thesis in 2019, and his M.S. in mechanics from Ecole Polytechnique in France.

"Meta-learning for remote sensing"

Abstract: In remotely sensed data, land cover types vary in appearance across geographic regions, yet also retain a shared set of characteristics everywhere on Earth. When performing classification, this raises the question of whether we should train one general model that can learn all global representations simultaneously, e.g. through pre-training, or develop many specialized models that each see data from one region.

This talk will highlight a different, "meta-learning" perspective that combines both strategies by training a model explicitly to adapt to new unseen regions. To accomplish this, we organize Earth observation data as a dataset of tasks, where classifying data from one region is considered a task, and explore model-agnostic meta-learning (MAML) for three common remote sensing applications: time series land cover classification, image land cover classification, and high resolution image segmentation. Following classic meta-learning approaches, we focus on a few-shot setting where only a small number of data points are available for an unseen region. Our results show that meta-learning approaches outperform traditional pretraining and fine-tuning schemes if the data representations of evaluated regions vary.

Bio: I’m a PhD student at Stanford studying applied math (ICME), advised by Professor David Lobell at the Center on Food Security and the Environment. I work on developing machine learning methods for remote sensing applications, especially in settings where ground truth labels are scarce. These methods are then applied to problems in sustainable agriculture and development, such as mapping where crops are grown in developing countries.