Investigating two super-resolution methods for downscaling precipitation: ESRGAN and CAR

Watson, Campbell D.; Wang, Chulin; Lynar, Timothy; Weldemariam, Komminist

doi:10.48550/arxiv.2012.01233

Cited by 6 publications

(6 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Machine learning approaches include Watson et al [2020], with the use of Generative Adversarial Networks (GANs) to increase the resolution of rainfall forecasts while simultaneously aiming to correct biases. This work was expanded upon by Price and Rasp [2022] by using a conditional GAN based on coarse weather variables.…”

Section: Alternative Probabilistic Forecasting Approaches For Rainfallmentioning

confidence: 99%

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

Huk

Adewoyin

Dutta

2023

Preprint

View full text Add to dashboard Cite

<p>This work develops a novel method for generating conditional probabilistic rainfall forecasts with temporal and spatial dependence. A two-step procedure is employed. Firstly, marginal location-specific distributions are modelled independently of one another. Secondly, a spatial dependency structure is learned in order to make these marginal distributions spatially coherent. <br />To learn marginal distributions over rainfall values, we propose a class of models termed Joint Generalised Neural Models (JGNMs). These models expand the linear part of generalised linear models with a deep neural network allowing them to take into account non-linear trends of the data while learning the parameters for a distribution over the outcome space.<br />In order to understand the spatial dependency structure of the data, a model based on censored copulas is presented. It is designed for the particularities of rainfall data and incorporates the spatial aspect into our approach. Uniting our two contributions, namely the JGNM and the Censored Spatial Copulas into a single model, we get a probabilistic model capable of generating possible scenarios on short to long-term timescales, able to be evaluated at any given location, seen or unseen. We show an application of it to a precipitation downscaling problem on a large UK rainfall dataset and compare it to existing methods.</p>

show abstract

Section: Alternative Probabilistic Forecasting Approaches For Rainfallmentioning

confidence: 99%

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

Huk

Adewoyin

Dutta

2023

Preprint

View full text Add to dashboard Cite

show abstract

“…These do not offer the black box benefits of ensembling strategies that automatically learn regression surfaces, requiring subject level knowledge on the number of basis functions and how to encode covariates. The generative adversarial networks that have recently been applied to spatial interpolation problems (Zhu et al, 2020;Manepalli et al, 2020;Watson et al, 2020a;Gao et al, 2020) similarly require some level of subject knowledge, with expected performance of deep learning networks depending on a large number of tuning structures.…”

Section: Parameter Tuningmentioning

confidence: 99%

Treeging

Watson,

Jerrett,

Reid

et al. 2021

Preprint

View full text Add to dashboard Cite

Treeging combines the flexible mean structure of regression trees with the covariancebased prediction strategy of kriging into the base learner of an ensemble prediction algorithm. In so doing, it combines the strengths of the two primary types of spatial and space-time prediction models: (1) models with flexible mean structures (often machine learning algorithms) that assume independently distributed data, and (2) kriging or Gaussian Process (GP) prediction models with rich covariance structures but simple mean structures. We investigate the predictive accuracy of treeging across a thorough and widely varied battery of spatial and space-time simulation scenarios, comparing it to ordinary kriging, random forest and ensembles of ordinary kriging base learners. Treeging performs well across the board, whereas kriging suffers when dependence is weak or in the presence of spurious covariates, and random forest suffers when the covariates are less informative. Treeging also outperforms these competitors in predicting atmospheric pollutants (ozone and PM 2.5 ) in several case studies. We examine sensitivity to tuning parameters (number of base learners and training data sampling proportion), finding they follow the familiar intuition of their random forest counterparts. We include a discussion of scaleability, noting that any covariance approximation techniques that expedite kriging (GP) may be similarly applied to expedite treeging.

show abstract

“…Generating model simulations of atmospheric processes at high spatial and temporal resolutions (super-resolution) have numerous applications including hybrid physical-model and machine-learning applications (Onishi et al, 2019), the dynamic downscaling of coarse resolution climate and weather information (Watson et al, 2020), and urban-climate feedback studies (Wu et al, 2021). Super-resolution modelling products (∆x < 100 m, ∆t < 1 s) can also provide desirable information at the scale of measurements during top-down campaigns which can be analyzed in conjunction with measurement data to: interpret observations, quantify uncertainty in the measurements, test the validity of assumptions in the employed top-down methodologies, and help fill the information gap in measurements.…”

Section: Introductionmentioning

confidence: 99%

“…To study these effects through model simulations, the model resolutions should be chosen to resolve dynamical processes (turbulence) at the spatio-temporal scales at which aircraft in-situ measurements are made. For instance, to simulate (and evaluate) in-situ measurements at a flying/sampling speed of 100 m/s (e.g., Conley et al, 2017;Gordon et al, 2015), the model should be able to simulate (and output) atmospheric fields at length and time scales of ∆x ≤ 100 m and ∆t ≤ 1 s. Recent real-case LES-modelling studies have commonly referred to such resolutions (∆x ≤ 250 m) as "super-resolution" (e.g., Wu et al, 2021;Onishi et al, 2019;Watson et al, 2020), herein we use the same terminology to describe our WRF model simulations.…”

Section: Introductionmentioning

confidence: 99%

Passive Tracer Modelling at Super-Resolution with WRF-ARW to Assess Mass-Balance Schemes

Fathi

Gordon²,

Chen³

2022

Preprint

View full text Add to dashboard Cite

Abstract. Accurate "super-resolution'' (Δx < 250 m) atmospheric modelling is useful for several different sectors (e.g., renewable energy, natural disaster prediction), and essential for numerous applications such as downscaling of weather and climate information to finer resolutions. It can also be used to interpret environmental observations during top-down retrieval campaigns by providing complementary data that closely correspond to real-world atmospheric pollution transport and dispersion conditions. In top-down retrievals (e.g., aircraft-based), errors in estimates can arise from assumptions about atmospheric dispersion conditions, uncertainties in measurements, and data processing. As discussed in this work and in our companion paper (Fathi and Gordon, 2022), super-resolution numerical model simulations can be utilized to investigate these sources of uncertainty and optimize the retrievals. In order to conduct a thorough model-based study of the atmospheric dynamical processes that can affect top-down retrievals, model simulations at super-resolutions on the scale of measurement frequency are required: sufficient to resolve the dynamical and turbulent processes at the scale at which measurements are conducted. Here, in the context of our modelling case studies with WRF, we demonstrate a series of best practices for improved (realistic) modelling of atmospheric pollutant dispersion at super-resolutions. These include careful considerations for grid quality over complex terrain, sub-grid TKE parameterization at the scale of large eddies, and ensuring local and global tracer mass-conservation. For this work, super-resolution (Δx ≤ 100 m, Δt ≤ 1 s) model simulations with Large-Eddy-Simulation sub-grid scale parameterization were developed and implemented using WRF-ARW. The objective was to resolve small dynamical processes inclusive of spatio-temporal scales of high-speed (e.g., 100 m/s) airborne measurements. This was achieved by down-scaling of reanalysis data from 31.25 km to 50 m through multi-domain model nesting in the horizontal and grid-refining in the vertical. Further, WRF dynamical-solver source code was modified to simulate passive-tracer emissions within the finest resolution domain. Different meteorological case studies and several tracer emission sources were considered. Model-generated fields were evaluated against observational data and also in terms of tracer mass-conservation. Results indicated model performance within 5 % of observational data in terms of sea level pressure, temperature and humidity, and agreement within one standard deviation between modelled and observed wind fields. Model performance in terms of tracer mass conservation was within 2 % to 5 % of model input emissions.

show abstract

Investigating two super-resolution methods for downscaling precipitation: ESRGAN and CAR

Cited by 6 publications

References 0 publications

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

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

Treeging

Passive Tracer Modelling at Super-Resolution with WRF-ARW to Assess Mass-Balance Schemes

Contact Info

Product

Resources

About