Temperature predicting by Stochastic Adversarial Video Prediction

In recent years, video prediction methods by deep learning have demonstrated success in computer vision applications, such as self-driving cars, human action prediction, and gesture recognition. The weather and climate communities are beginning to investigate the use of these advanced methods in the context of weather and climate forecasting (Reichstein, M., et al., 2019). Current weather forecasting systems rely on numerical weather prediction models which divide the earth system into grid boxes and numerically solve complex differential equations describing the temporal evolution of the system in terms of momentum, energy and mass. While these models have reached remarkable quality in many aspects, high resolution simulations in space and time require huge computational resources (Zängl, Günther, et al. 2015). In addition to errors inherent from observational data, which are used for the initialization, discretization errors and the need of parametrization schemes for a variety of physical processes diminish the accuracy of numerical weather predictions. Two specific examples in this context are clouds and precipitation. Deep learning models promise to discover non-linear spatio-temporal properties from heterogeneous weather observations in a data-driven way and are, once these models have been trained, computationally much cheaper.

There are certain similarities between the deep learning tasks of video prediction and weather forecasting. Among others, they both explore spatio-temporal patterns from previously observed data to generate (i.e. forecast) the future frames. This work, then, explores the generalization capability and adaptation of the video prediction method and uses a GAN-based architecture, such as the stochastic adversarial video prediction (SAVP) (Figure 7) when applied to a weather dataset containing temperature, pressure, and geopotential, including a comparison of transfer learning and end-to-end training.

Figure 8: SAVP Architecture adopted from (Lee, A.X., et al., 2018)

Lee, A.X., Zhang, R., Ebert, F., Abbeel, P., Finn, C., Levine, S.: Stochastic adversarial video prediction. arXiv preprint arXiv:1804.01523 (2018)
Reichstein, M., Camps-Valls, G., Stevens, B., Jung, M., Denzler, J., Carvalhais, N., et al.: Deep learning and process understanding for data-driven earth system science. Nature 566(7743) (2019) 195–204
Zängl, Günther, et al. “The ICON (ICOsahedral Non‐hydrostatic) modelling framework of DWD and MPI‐M: Description of the non‐hydrostatic dynamical core.” Quarterly Journal of the Royal Meteorological Society 141.687 (2015): 563-579.

Near surface ozone predictions

Human beings, animals and crops are directly exposed to the ambient airmass of the lower part of the atmosphere (planetary boundary layer). This airmass may contain several pollutants like ozone which is harmful to living beings (WHO, 2013; Bell et al., 2014; Lefohn et al., 2017; Fleming et al., 2018 ) and certain crops (Avnery et al., 2011; Mills et al., 2018). Therefore, the prediction of ozone concentrations is of significant importance to issue warnings for the public if high ozone concentrations are foreseeable.

We develop deep neural networks to predict near-surface ozone concentrations for more than 300 German measurement sites. Our recent model consists of multiple convolutional layers which we group into individual inception blocks (extension of Szegedy, 2015). Each inception block consists of three convolutional, and two pooling towers (max, mean) [Figure 6]. We use chemical precursors and meteorological variables of the seven previous days as input to predict the daily maximum eight-hour average (dma8eu) for a lead time of up to four days.

We create a workflow which ensures that the full workflow, including downloading the data, preprocessing, training, and postprocessing steps, is reproducible.

Figure 7:Inception block with dimensionality reduction and batch normalisation before activation. Successive layers are grouped to towers. The input and output layer of an inception block are colour coded in green, pooling layers in red, filter reduction layers

WHO, 2013: http://www.euro.who.int/en/health-topics/environment-and-health/air-quality/publications/2013/health-risks-of-air-pollution-in-europe-hrapie-project.-recommendations-for-concentrationresponse-functions-for-costbenefit-analysis-of-particulate-matter,-ozone-and-nitrogen-dioxide
Bell et al., Who is More Affected by Ozone Pollution? A Systematic Review and Meta-Analysis, 2014: https://doi.org/10.1093/aje/kwu115
Lefohn et al., Responses of human health and vegetation exposure metrics to changes in ozone concentration distributions in the European Union, United States, and China, 2017: https://doi.org/10.1016/j.atmosenv.2016.12.025
Fleming et al., Tropospheric Ozone Assessment Report: Present-day ozone distribution and trends relevant to human health, 2018: https://doi.org/10.1525/elementa.273
Avnery et al., Global crop yield reductions due to surface ozone exposure: 1. Year 2000 crop production losses and economic damage, 2011; https://doi.org/10.1016/j.atmosenv.2010.11.045
Mills et al., Ozone pollution will compromise efforts to increase global wheat production, 2018: https://doi.org/10.1111/gcb.14157
Szegedy et al., Going deeper with convolutions, 2015: https://doi.org/10.1109/CVPR.2015.7298594