Convolutional conditional neural processes for local climate downscaling

Abstract

Abstract. A new model is presented for multisite statistical downscaling of temperature and precipitation using convolutional conditional neural processes (convCNPs). ConvCNPs are a recently developed class of models that allow deep learning techniques to be applied to off-the-grid spatio-temporal data. This model has a substantial advantage over existing downscaling methods in that the trained model can be used to generate multisite predictions at an arbitrary set of locations, regardless of the availability of training data. The convCNP model is shown to outperform an ensemble of existing downscaling techniques over Europe for both temperature and precipitation taken from the VALUE intercomparison project. The model also outperforms an approach that uses Gaussian processes to interpolate single-site downscaling models at unseen locations. Importantly, substantial improvement is seen in the representation of extreme precipitation events. These results indicate that the convCNP is a robust downscaling model suitable for generating localised projections for use in climate impact studies, and motivates further research into applications of deep learning techniques in statistical downscaling.

CDAnet: A Physics‐Informed Deep Neural Network for Downscaling Fluid Flows

“…Specifically, purely data-driven approaches based on super-resolution were applied for enhancing the resolution of coarse information without constraining the solution by the physics (e.g. Vaughan et al (2021)). Similar work was introduced for data compression, where the training and validation of the network were conducted using the same dataset, comprising a single or multiple solution trajectories (Jiang et al, 2020a).…”

Joint Generalized Neural Models and Censored Spatial Copulas for Probabilistic Rainfall Forecasting

“…Remark: It is worth reiterating that our data does not include precipitation weather forecasts as a predictor, unlike the above-mentioned studies of Harris et al [2022], Vaughan et al [2021]. Consequently, the performance of the benchmark methods obtained in their respective articles cannot be directly compared with the performances presented in this study (of the benchmarks as well as our approach).…”

“…We begin by comparing our method (denoted Cens-JGNM) to two competing benchmark methods: the variational auto-encoder generative adversarial network (VAE-GAN) [Harris et al, 2022] and the convolutional conditional neural process (ConvCNP) [Vaughan et al, 2021]. Both these methods as well as our own were fitted to the grided data detailed in Section 2.1 spanning a 20-year period, divided into a training set from 1979 to 1993 and a validation set from 1993 to 1999.…”

“…Further, our model can be evaluated on new locations after training, unlike most of these methods. Vaughan et al [2021] employ a convolutional conditional neural process model with parametric marginals, explicit modelling of zero values and inclusion of out-of-sample locations. Their approach is very similar to the JGNM, although their model fails at capturing spatial dependence to the same extent as our censored copula method, as evidenced in Section 5.1.…”

Environmental sensor placement with convolutional Gaussian neural processes

“…Convolutional neural processes (ConvNPs) are a recent class of ML models that have shown promise in modelling environmental variables. For example, ConvNPs can outperform a large ensemble of climate downscaling approaches (Vaughan et al, 2021;Markou et al, 2022) and integrate data of gridded and point-based modalities (Bruinsma et al, 2023). One variant, the convolutional Gaussian neural process (ConvGNP; Bruinsma et al, 2021;Markou et al, 2022), uses neural networks to parameterise a joint Gaussian distribution at target locations, allowing them to scale linearly with dataset size while learning mean and covariance functions directly from the data.…”