CliGAN: A Structurally Sensitive Convolutional Neural Network Model for Statistical Downscaling of Precipitation from Multi-Model Ensembles

Abstract

Despite numerous studies in statistical downscaling methodologies, there remains a lack of methods that can downscale from precipitation modeled in global climate models to regional level high resolution gridded precipitation. This paper reports a novel downscaling method using a Generative Adversarial Network (GAN), CliGAN, which can downscale large-scale annual maximum precipitation given by simulation of multiple atmosphere-ocean global climate models (AOGCM) from Coupled Model Inter-comparison Project 6 (CMIP6) to regional-level gridded annual maximum precipitation data. This framework utilizes a convolution encoder-dense decoder network to create a generative network and a similar network to create a critic network. The model is trained using an adversarial training approach. The critic uses the Wasserstein distance loss function and the generator is trained using a combination of adversarial loss Wasserstein distance, structural loss with the multi-scale structural similarity index (MSSIM), and content loss with the Nash-Sutcliff Model Efficiency (NS). The MSSIM index allowed us to gain insight into the model’s regional characteristics and shows that relying exclusively on point-based error functions, widely used in statistical downscaling, may not be enough to reliably simulate regional precipitation characteristics. Further use of structural loss functions within CNN-based downscaling methods may lead to higher quality downscaled climate model products.

DL4DS -- Deep Learning for empirical DownScaling

“…The SSIM is a computer vision metric used for measuring the similarity between two images, based on three comparison measurements: luminance, contrast and structure. The MS-DSSIM loss has been used by Chaudhuri and Robertson (2020) for the task of statistical downscaling.…”

“…Through iterative adversarial training, the resulting generator can produce outputs consistent with the distribution of real data, while the discriminator cannot distinguish between the generated high-resolution data and the ground truth. Adversarial training has been used in DL-based downscaling approaches proposed by Leinonen et al (2020), Stengel et al (2020) and Chaudhuri and Robertson (2020). In out tests, the generated high-resolution fields produced by the CGAN generators exhibit moderate variance, even when using the Monte Carlo dropout technique (Gal and Ghahramani, 2016) which amounts to applying dropout at inference time.…”

Machine Learning Modeling of Climate Variability Impact on River Runoff

“…Among numerous types of NNs used in water science, the convolution neural networks (CNN) architectures have become increasingly popular in recent years, perhaps due to their impressive record of achievements in image analysis, then corroborated for various types of time series (e.g., speech recognition and synthesis), and due to the quickly growing repertoire of extensions (e.g., batch normalization, dropout, modern self-tuning training algorithms) and software libraries that made them effective at larger scales (more complex, deeper models, and efficient learning from large data volumes) and deployable on highly efficient hardware architectures (GPUs and dedicated platforms). Only within the last two years have CNNs been used for estimation of discharge for an ungauged watershed [70]; for estimation of pollutant loads (biochemical oxygen demand and total phosphorus) in ungauged watersheds [71]; for statistical downscaling of precipitation from multi-model ensembles [72]; and, in combination with the popular variant of recurrent NN, long short-term memory (LSTM), to simulate water quality (including total nitrogen, total phosphorus, and total organic carbon) in Korea [73]. This list is by no means exhaustive.…”

Climate Change Multi-Model Projections in CMIP6 Scenarios in Central Hokkaido, Japan

“…For example, Kreienkamp et al (2020) downscaled CMIP6 GCM outputs in Germany using the statistical-empirical downscaling approach 31 . Chaudhuri and Robertson (2020) developed the deep neural network model with a structural sensitivity to downscale large-scale annual maximum precipitation from 9 CMIP6 GCMs in Great Bear Lake in Northwest Territories, Canada 32 . However, few studies on SDSM of temperature and precipitation based on CMIP6 outputs have been undertaken.…”