Search

This paper aims to explore machine learning techniques for post‐processing high‐resolution Numerical Weather Prediction (NWP) products for the early detection of convection. Data from the Arome Ensemble Prediction System and satellite observations from the Rapidly Developing Thunderstorm (RDT) product by Météo‐France are used to train a recurrent neural network model to predict areas of total convection and moderate convection. The learning task is formulated as a binary classification problem using a long short‐term memory (LSTM) network architecture. Results from the LSTM model are compared with an object‐based probabilistic approach to forecast convection using metrics such as a receiver operating characteristics (ROC) curve, the Brier score and reliability. Results indicate that the LSTM model performs similarly to the object‐based probabilistic benchmark when classifying moderate convection areas and shows improved skill when classifying areas of total convective. Finally, the LSTM model results are presented within an air traffic management context to showcase the potential use of machine learning models within an operational application. [ABSTRACT FROM AUTHOR]

A global sensitivity analysis of the convective-scale Application of Research to Operations at Mesoscale (AROME)model is performed in order to determine the most influential parameters on the forecast of different near-surface variables. For that purpose, the Morris method is applied to 21 parameters from six different physical and dynamical parametrization schemes, over different seasons. Results highlight a set of eight parameters with a noticeable influence on most variables, in particular 10m wind speed and precipitation forecasts. The sensitivity of parameter uncertainties is also examined on different spatio-temporal scales. A clear diurnal cycle of parameters influence is observed in summer, in close connectionwith the convective activity. In addition, the spatial distribution of parameters influence is mostly consistent with the underlying distribution of weather forecasts. A Sobol’ sensitivity analysis, based on surrogate models, mostly confirms Morris conclusions and highlights some interactions between parameters. [ABSTRACT FROM AUTHOR]

Object‐based methods are commonly used for verification and postprocessing of high‐resolution precipitation forecasts. They usually detect objects based on intensity criteria only, without considering the spatial organization of rainfall, known as texture. This article evaluates the performance of several machine‐learning methods to detect "continuous" and "intermittent" rainfall patterns in the forecasts of the French convective‐scale Arome model. A sliding‐window segmentation algorithm, which applies a classification model at each grid point, is implemented. Several classifiers and input textural features are compared. Overall, intermittent precipitation is the most difficult to detect. The random forest classifier is shown to provide the best classification results independently of the predictor used, with a surprising ability to extract a relevant signal from a synthetic descriptor such as the rainfall cumulative distribution function, as well as from the large amount of unprocessed information provided by neighbouring grid points. On the other hand, the logistic regression classifier needs a texture‐oriented predictor, such as the statistics derived from the grey‐level co‐occurrence matrix, to perform well. Global insight into model behaviour is then obtained by examining the importance of input features. Finally, we show that random forests trained on Arome deterministic forecasts can be applied successfully to discriminate between precipitation textures in different Arome configuration outputs and gridded observations. [ABSTRACT FROM AUTHOR]

In order to contribute to ongoing efforts on tropical cyclone (TC) forecasting, a new, convection‐permitting, limited‐area coupled model called AROME‐Indian Ocean (AROME‐IO) was deployed in the Southwest Indian Ocean basin (SWIO) in April 2016. The skill of this numerical weather predicting system for TC prediction is evaluated against its coupling model (European Center for Medium Range Weather Forecasting‐Integrated Forecasting System [ECMWF‐IFS]) using 120‐hr reforecasts of 11 major storms that developed in this area over TC seasons 2017–2018 and 2018–2019. Results show that AROME‐IO generally provides significantly better performance than IFS for intensity (maximum wind) and structure (wind extensions, radius of maximum wind) forecasts at all lead times, with similar performance in terms of trajectories. The performance of a prototype, 12‐member ensemble prediction system (EPS), of AROME‐IO is also evaluated on the case of TC Fakir (April 2018), a storm characterized by an extremely low predictability in global deterministic and ensemble models. AROME‐IO EPS is shown to significantly improve the predictability of the system with two scenarios being produced: a most probable one (~66%), which follows the prediction of AROME‐IO, and a second one (~33%) that closely matches reality. [ABSTRACT FROM AUTHOR]

Traditional pointwise verification scores are not always appropriate for the evaluation of high‐resolution precipitation forecasts because of double‐penalty problems. An alternative approach, based on the identification of homogeneous rainfall areas called "precipitating objects", allows forecast evaluation at a larger and thus more predictable scale, and specific information about the nature of errors (e.g. location, size, intensity) can be obtained. A novel object detection method is first introduced and the object‐based verification of precipitation forecasts from the convective‐scale deterministic and ensemble models Arome and Arome‐EPS is then discussed, using several scores and diagnostics. Three types of precipitating objects characterizing total, moderate and heavy rainfall are considered. In the second part, object‐based metrics are used to compute objective weights for time‐lagged ensemble forecasts, based on their performance at early forecast ranges. The weights obtained clearly depend on the meteorological situation and on the precipitation type, reflecting for instance the lower predictability of moderate precipitation compared to total precipitation. There is also a dependence on the production time with, on average, slightly larger and more homogeneous weights associated with the most recent run. However, in some situations of moderate and heavy rainfall, a relevant signal can be extracted from older runs. It is finally shown that object‐based weights are better suited than classical quadratic weights to improve nowcasting performance. [ABSTRACT FROM AUTHOR]

Limited‐area ensemble predictions can be sensitive to the specification of lateral boundary conditions, which are often built by subsampling larger ensembles. Using the operational PEARP and AROME‐EPS ensembles, we compare several subsampling methods, including random selection, representative members, and a new selection method. The tests show that the algorithms used for the clustering and the member selection have a significant impact on the resulting ensembles. Clustering‐based methods are shown to outperform random subsampling, mostly (but not only) because they change the ensemble spread. Cluster sizes can be highly variable, which can complicate ensemble interpretation. We present a subsampling algorithm that has little impact on performance scores, but better preserves ensemble spread and produces nearly equally likely members by limiting cluster size variability. Limited‐area ensemble predictions can be sensitive to the specification of lateral boundary conditions, which are often built by subsampling larger ensembles. Using the operational PEARP and AROME‐EPS ensembles, we compare several subsampling methods, including random selection, representative members, and a new selection method. [ABSTRACT FROM AUTHOR]

The relative benefits of ensemble size and model resolution are investigated within the AROME-France convective-scale ensemble prediction system (EPS), which operationally runs 12 perturbed members at 2.5 km horizontal resolution. This baseline configuration is compared with two auxiliary ensemble experiments, which are run at resolutions of 2.5 and 1.3 km, with 34 and 12 members respectively. In addition, post-processing techniques including neighbourhood approaches and time-lagging are examined as potential alternatives to increase the sample size at a lower computational cost. Probabilistic verification of the raw EPS outputs indicates that increasing model resolution is more beneficial at very short ranges, whereas increasing the ensemble size has a larger impact at longer forecast ranges, as predictability decreases and more members are required to sample the larger uncertainty better. The neighbourhood processing confirms these conclusions and is shown to improve precipitation forecasts of all EPS configurations significantly, especially the smaller-size ensembles. Hence, it can be considered as a viable substitute to running additional AROME members for the first 24 h of forecasts. Time-lagging of three successive ensemble productions also appears to be a competitive approach to improve the ensemble forecast skill, particularly at very short ranges, where AROME-EPS is known to be underdispersive. The performance of the time-lagged 36 member ensemble is generally close to or even better than the performance of the single 34 member ensemble for the whole forecast range and for different surface weather variables. Overall, given the relative costs and skills of the different EPS configurations, we suggest that resources should primarily be spent on increasing ensemble size, by combining both additional members and post-processing methods. In addition, the performance of the 1.3 km ensemble should be examined further for explicit high-impact weather and with an appropriate tuning of physics and surface perturbations. [ABSTRACT FROM AUTHOR]

In order to cope with small-scale unpredictable details of mesoscale structures in cloud-resolving models, it is suggested that model outputs are processed following a fuzzy object-oriented approach to extract and track precipitating features (which are associated with a higher predictability than the direct model outputs). The present approach uses the particle filter method to recognize patterns based on predefined texture or spatial variability of the model output. This provides an ensemble of precipitating objects, which are then propagated in time using a stochastic advection-diffusion process. This method is applied to both deterministic and ensemble forecasts provided by the AROME-France convective-scale model. Specific case-studies support the ability of the approach to handle precipitation of different types. [ABSTRACT FROM AUTHOR]

The AROME-EPS convection-permitting ensemble prediction system has been evaluated over the HyMeX-SOP1 period. Objective verification scores are computed using dense observing networks prepared for the HyMeX experiment. In probabilistic terms, the AROME-EPS ensemble performs better than the AROME-France deterministic prediction system, and a state-of-the-art ensemble at a lower resolution. The strengths and weaknesses of AROME-EPS are discussed. Here, impact experiments are used to study perturbation schemes for the initial conditions and the model surface. Both have a significant effect on the ensemble performance. The interactions between the perturbations of lateral boundaries, initial conditions and surface perturbations are studied. The consistency between initial and lateral perturbations is found to be unimportant from a meteorological point of view. Ensemble data assimilation is not as effective as a simpler surface perturbation scheme, and it is noted that both approaches could be usefully combined. [ABSTRACT FROM AUTHOR]

Convective-scale ensemble prediction systems (EPSs) are often initialized with downscaled initial condition perturbations (ICPs) from a global coarser EPS. Although downscaled ICPs have been shown to have a positive impact at short ranges, they cannot represent the uncertainty at small scales. Hence, there is a spin-up of around 9-12 h until the forecast perturbations develop realistic small-scale structures. On the other hand, ensemble data assimilation (EDA) is a common approach to obtain initial perturbations at all scales resolved by the numerical model. However, the high computational cost of EDA systems severely limits their size and their resolution. An alternative cheaper method to derive small-scale ICPs is considered here, based on a random sampling of themodel backgrounderror covariances. This article provides an evaluation of random and EDA-based IC perturbation methods against the baseline downscaling approach, in the framework of the pre-operational convective-scale EPS developed at Météo-France with the AROME-France model at a 2.5 km horizontal resolution. Small-scale IC perturbation methods are shown to significantly improve the short-range EPS performance for surface weather variables. For 2m temperature and 10m wind speed, random ICPs give as good results as the EDA, owing to the very short spin-up of random perturbations. Precipitation forecasts are also strongly improved during the first six forecast hours, especially when initial humidity perturbations are included. The sensitivity of the EPS performance to the EDA size, horizontal resolution and representation of model errors is discussed. It is found that a large fraction of the initial uncertainty can be properly described with an EDA of reasonable size and at a slightly coarser resolution than the EPS. [ABSTRACT FROM AUTHOR]

Abstract: Background‐error variances estimated from a small‐size ensemble of data assimilations need to be filtered because of the associated sampling noise. Previous studies showed that objective spectral filtering is efficient in reducing this noise, while preserving relevant features to a large extent. However, since such filters are homogeneous, they tend to smooth small‐scale structures of interest. In many applications, nonlinear thresholding of wavelet coefficients has proved to be an efficient technique for denoising signals. This algorithm proceeds by thresholding the wavelet coefficients of the noisy signal using an estimated threshold. This is equivalent to applying an adaptive local spatial filtering. A quasi‐optimal value for the threshold can be computed from the noise variance. We show that the statistical properties of the sampling noise associated with the estimation of background‐error variances can be used to calculate the noise level and the appropriate threshold value. This method is first applied to 1D academic examples, with emphasis on correlated and heterogeneous noises. This approach is shown to outperform the commonly used homogeneous filters, since it automatically adapts to the local structure of the signal. We also show that this technique compares favourably to a heterogeneous diffusion‐based filter, with the advantage of requiring less trial‐and‐error tuning. These results are next confirmed in a more realistic 2D problem, using the Arome‐France convective‐scale model. [ABSTRACT FROM AUTHOR]

Background-error length-scales are key components of the correlation model since they give an idea of the distance over which the influence of an observation extends. The length-scales can be estimated from an ensemble of perturbed forecasts, however the finite size of the ensembles in operational applications introduces a relatively large sampling noise on the estimated length-scales. We first give some insight into the statistical properties of the noise. On the one hand, the noise variance can be analytically expressed as a simple function of the length-scale and the ensemble size. On the other hand, it is shown that the spatial structure of the noise is relatively small-scale and is directly related to the heterogeneity of background error. Such spatial properties tend to support the use of local averaging techniques to remove the noise. This is explored with the application of homogeneous and heterogeneous averagings based on the integration of the diffusion equation. It is observed that both filters improve the accuracy of the length-scale estimates, with the heterogeneous filter being more efficient in the regions where the length-scales are small and rapidly varying. Copyright © 2012 Royal Meteorological Society [ABSTRACT FROM AUTHOR]

Background-error variances estimated from a finite size ensemble of data assimilations are affected by sampling noise, which degrades the accuracy of the variance estimates. Previous work highlighted the close link between the spatial structures of background error and the associated sampling noise, and demonstrated the ability of local spatial averaging to remove this sampling noise. Existing filtering techniques commonly assume a homogeneous smoothing of the estimated variances. However, this assumption can be inadequate to represent error structures of varying scales, e.g. small-scale errors associated with localized severe weather events. To answer this problem, this article introduces and examines a heterogeneous filter based on the knowledge of the local spatial properties of the sampling noise. The filtering is realized with a diffusion process, and the diffusion coefficient is parametrized according to the local correlation length-scale of the sampling noise. This enables the diffusion coefficient to vary spatially in such a way as to encourage smoothing in regions where the background error is large scale in preference to regions where the error is small scale. A simulated 1D framework is considered to test the proposed approach. It is shown that the filtering using a spatially varying diffusion coefficient is able to preserve high-frequency variance structures, while this information tends to be smoothed with homogeneous filtering. The benefits of applying heterogeneous filtering are particularly pronounced with small ensemble sizes and in the vicinity of localized variance maxima. Copyright © 2012 Royal Meteorological Society [ABSTRACT FROM AUTHOR]

Ensemble prediction systems at the convective scale are often under-dispersive. In order to alleviate this problem, a time-lagged ensemble can be created from ensemble forecasts initialized at different production times. While an equal-weight combination of lagged forecasts generally provides competitive results, this article introduces and discusses the efficiency of an objective weighting. The proposed approach is based on nonlinear Bayesian filtering, and the weights are determined online for each member according to the observation likelihood. The method is illustrated with short-range ensemble forecasts provided by the cloud-resolving AROME-France model. A time-lagged ensemble is then constructed from the current ensemble forecasts combined with older ensemble forecasts started 6 and 12 h earlier. It is first shown that the weighting scheme provides reasonable results, in particular it is able to detect differences in forecast quality due to different production times. The question whether these unequal flow-dependent weights can be successfully applied to the members of the time-lagged ensemble is then examined. Results indicate that the weighting does not lead to a noticeable gain in forecast quality. Possible reasons for this limited impact are discussed. [ABSTRACT FROM AUTHOR]

In order to contribute to ongoing efforts on tropical cyclone (TC) forecasting, a new, convection‐permitting, limited‐area coupled model called AROME‐Indian Ocean (AROME‐IO) was deployed in the Southwest Indian Ocean basin (SWIO) in April 2016. The skill of this numerical weather predicting system for TC prediction is evaluated against its coupling model (European Center for Medium Range Weather Forecasting‐Integrated Forecasting System [ECMWF‐IFS]) using 120‐hr reforecasts of 11 major storms that developed in this area over TC seasons 2017–2018 and 2018–2019. Results show that AROME‐IO generally provides significantly better performance than IFS for intensity (maximum wind) and structure (wind extensions, radius of maximum wind) forecasts at all lead times, with similar performance in terms of trajectories. The performance of a prototype, 12‐member ensemble prediction system (EPS), of AROME‐IO is also evaluated on the case of TC Fakir (April 2018), a storm characterized by an extremely low predictability in global deterministic and ensemble models. AROME‐IO EPS is shown to significantly improve the predictability of the system with two scenarios being produced: a most probable one (~66%), which follows the prediction of AROME‐IO, and a second one (~33%) that closely matches reality. [ABSTRACT FROM AUTHOR]