Generation of atmospheric cloud fields using generative adversarial networks

Abstract

Cloud fields have a great impact on the weather and climate of the Earth. Modern climate models that include feedback from cloud fields present large uncertainties in their predictions. To improve models of climate systems large amounts of high-quality data of cloud fields are necessary. The Cloud Profiling Radar (CPR) on the CloudSat satellite provides high-quality vertical data of cloud fields but is limited in its geographical coverage. The Moderate Resolution Imaging Spectroradiometer (MODIS) on the Aqua satellite provides data that only discerns the top layer of clouds but has far greater spatial coverage. This project implements two generative adversarial networks (GAN), with the intention of extending the current available CPR data set. One network generates vertical cloud data using random noise as input. The other network uses MODIS data as input to generate vertical cloud data tied to a specific geolocation, this network is accordingly a conditional generative adversarial network (CGAN). The two neural networks are compared to a common method used for generating synthetic cloud fields, the Iterative Amplitude Adjusted Fourier Transform (IAAFT). The methods are compared in regard to multiple different statistical and physical properties, including ice water path, cloud-top height, and spatial autocorrelation. The results of both GAN are promising with regard to generating realistic cloud fields. Individual generated cloud scenes from GAN and CGAN show a stronger visual resemblance to real radar scenes than scenes generated with IAAFT. The three different methods vary in performance when analysed based on the statistics of their output. For example, the IAAFT method outperforms both the GAN and the CGAN when it comes to recreating the large scale vertical cloud distributions, but shows a clear weakness when it comes to capturing the internal structure of cloud fields, measured by the autocorrelation. A reoccurring problem with the GAN is the overfitting of the training data and mode collapse. This aspect needs further improvement through re-training of the networks with a larger data set and possibly uneven weighting between the generator and discriminator. After the inclusion of such improvements, the two GAN developed in this project are expected to show promising results for including generated radar data in applications.