Statistical downscaling of daily temperature and precipitation over China using deep learning neural models: Localization and comparison with other methods

Abstract

Convolutional neural network (CNN) is an effective tool for extracting interpretable information from big data and has been recently used as a promising approach for statistical downscaling. In this study, CNN models of different configurations are used to downscale daily temperature and precipitation over China with the use of large‐scale atmospheric variables from ECMWF Interim reanalysis (ERI) and high‐resolution gridded observations as predictors and predictands respectively. A 21‐year period from 1979 to 1999 is used for calibration and a relatively warmer period during 2000–2017 is used for validation, which helps to examine the extrapolation capability of models. It is shown that model performance varies among different configurations. For a realistic multi‐site downscaling over whole China, the convolutional process is indispensable and much more spatial features are required to parameterize temperature characteristics than precipitation. As compared with ERI, CNN model shows added value in reproducing geographic distributions of seasonal mean climate and seasonal cycle as well as reducing biases in mean and extreme percentiles for both temperature and precipitation. However, ERI performs better in terms of temporal correlations. Then the model is further compared with Generalized Linear regression Model (GLM) and two quantile mapping based techniques including bias correction and spatial disaggregation (BCSD) and bias correction and climate imprint (BCCI). It is shown that bias correction methods show superior performances to other models in reducing biases and representing temporal correlations especially for precipitation. CNN model achieves better precipitation downscaling performances than GLM. It also achieves good skills in reproducing seasonal cycle of temperature and frequency distributions of daily precipitation, and presents better stabilities between the calibration and validation period. These results indicate that CNN model has good potential for downscaling application over large regions (e.g., continents).

Easy-to-use spatial random-forest-based downscaling-calibration method for producing precipitation data with high resolution and high accuracy

“…Vegetation types have a significant impact on fluxes of sensible and latent heat into the atmosphere, apparently influencing the humidity of the lower atmosphere and further affecting moist convection (Spracklen et al, 2012). Therefore, as an indicator of vegetation activity, NDVI has been widely adopted to estimate precipitation (Wu et al, 2019;Immerzeel et al, 2009).…”

“…Nevertheless, these methods are based on some strict assumptions, which might not be satisfied in reality (Zhang et al, 2021;Wu et al, 2020). To this end, ML-based calibration methods have been widely used, such as quantile regression forest (QRF) (Bhuiyan et al, 2018), ANN (Yang and Luo, 2014;Pham et al, 2020), deep neural network (Tao et al, 2016), RF (Baez-Villanueva et al, 2020), convolutional neural network (CNN) (Wu et al, 2020), SVM and extreme learning machine (Zhang et al, 2021).…”

On the suitability of deep convolutional neural networks for continental-wide downscaling of climate change projections

“…To do so, they used the experimental framework defined in VALUE (Experiment 1) to compare the results provided by CNNs with those obtained from a set of other more classical, standard techniques i.e., generalized linear models, concluding that CNNs are well suited for continental-wide applications. There have been similar studies over North America (Pan et al 2019) and China (Sun and Lan 2020), all showing that CNNs achieve similar or better performance than standard SDMs. Moreover, CNNs circumvent the problem of feature selection/extraction-which is highly case-dependent and becomes a very complex task to accomplish in classical downscaling methods-by performing an implicit manipulation of the input space in the internal structure of the network (Baño-Medina 2020).…”

Deep Learning for Daily Precipitation and Temperature Downscaling

“…Recent studies have evaluated the performance of simple plain CNN for downscaling daily temperature and precipitation relative to classic statistical downscaling methods and the results show limited improvement (Baño‐Medina et al., 2020; Miao et al., 2019; Pan et al., 2019; L. Sun & Lan, 2020; Vandal, Kodra, Ganguly, et al., 2018). For example, Baño‐Medina et al.…”