The rapid expansion of the French offshore wind sector requires a critical reassessment of structural durability in the face of evolving marine conditions driven by climate change. Traditional design methodologies assuming stationary conditions are inadequate for climate-resilient infrastructure, as they fail to account for evolving wave patterns, storm tracks, and extreme event frequencies (Amlashi, 2024; Barkanov et al., 2024).This study introduces a non-stationary GEV framework to quantify future changes in significant wave height using monthly maxima, reducing uncertainty in design tools for offshore wind farms under climate change. While non-stationary GEV models are well established, their application to offshore wind design using climate projections remains limited. The main objective is to derive an equivalent stationary design level that accounts for time-varying extremes over the lifetime of structures. Based on CMIP6 climate models and reanalysis data, results reveal a projected trend towards a more pronounced seasonal contrast along the French Atlantic and English Channel coasts under future scenarios (SSP1-2.6 and SSP5-8.5), whereas the French Mediterranean Sea exhibits weaker increases in extremes and larger uncertainties (inter-model spread). Projections indicate more intense winters and calmer summers, along with a shift in the seasonal cycle. Overall, the multi-model ensemble suggests an increase in the design levels for extreme sea states.The research concludes by defining a new methodology for calculating an equivalent design level over the structure's operational lifespan. This tool is deemed essential for ensuring the resilience and economic viability of future offshore wind farms in a changing climate.
Recurrent flooding in the Marahoué region, particularly in Bouaflé and Zuénoula, underscores the need for reliable and operational tools to anticipate hydrological risks and support early warning systems. This study presents a rainfall forecasting framework based on Long Short-Term Memory (LSTM) neural networks, integrating satellite-derived precipitation products and reanalysis-based atmospheric variables to predict daily rainfall at the Bouaflé and Zuénoula stations at t + 1, t + 3, and t + 7 d lead times. The performance of the LSTM model was systematically evaluated and compared with commonly used reference models, namely Random Forest (RF), Extra Trees (ET), and XGBoost (XGB), using standard statistical metrics (R2, NSE, Pearson correlation coefficient R, normalized RMSE, and MAE). The results show that the LSTM model consistently outperforms the reference models across all forecasting horizons and at both study stations. At short and medium lead times (t + 1 and t + 3), LSTM exhibits strong predictive skill, with R2
Precipitation generators are statistical models used to produce synthetic sequences of (multisite) precipitation for hydrological applications such as flood risk assessment and water resource management. Among these, approaches based on generalized linear models (GLMs) are widely used and often perform competitively with state-of-the-art alternatives, but they face limitations in representing seasonal variation in extremes and in flexibly capturing covariate effects on the precipitation distribution. In this paper, we extend the GLM framework in two directions. First, we introduce generalised additive models for location, scale and shape (GAMLSS) for precipitation generation. These models allow the use of spline-based model terms to flexibly capture covariate effects and allow covariates to influence multiple distributional parameters, thereby increasing flexibility in representing variation in both the mean and variance of the precipitation distribution. Second, we adapt a transformed Gaussian fields approach to jointly account for spatial dependence in both precipitation occurrence and intensity, thus allowing for potential cross-dependence between the two. A further contribution is to investigate the sensitivity of model performance to data resolution, highlighting that rounding in data pre-processing can substantially affect the reproduction of extremes. Using a well-studied daily precipitation dataset, we demonstrate that these extensions improve the realism of simulated sequences, particularly with respect to extremes, and capture spatial dependence well in both occurrence and intensity.
We investigate the influence of time-varying meteoceanic conditions on coastal flooding under the prism of rare events. Focusing on conditions observed over half tidal cycles, we observe that such data fall within the framework of functional extreme value theory, but violate standard assumptions due to temporal dependence and short-tailed behavior. To address this, we propose a two-stage methodology. First, we introduce an autoregressive model to reduce temporal dependence between cycles. Second, considering the model residuals, we adapt existing techniques based on Pareto processes. This allows us to build a simulator of extreme scenarios, by applying inverse transformations. These simulations depend on an initial time series, which can be randomly selected to tune the desired level of extremes. We validate the simulator performance by comparing simulated time series with observations, through several criteria, based on principal component analysis, extreme value analysis, and classification algorithms. The approach is applied to the surge data, on the Gâvres site, located in southern Brittany, France.
Burn-scar persistence in satellite imagery is treated as a lifetime process and modelled within a doubly censored Weibull framework arising from the finite temporal extent of satellite archives. Episodes may appear doubly censored in calendar time because the Landsat record spans a limited analysis window, so that for some pixels neither the true ignition time nor the full recovery are directly observed. On the persistence-time scale used for inference, incomplete episodes contribute as right-censored lifetimes within a unified likelihood formulation. A robust weighted-likelihood estimator with Huber-type influence weights downweights atypically long-lived scars without excluding any observations, thereby limiting the impact of outliers while preserving the central structure of the persistence distribution. The framework is applied to NDVI-based burn-scar persistence derived from Landsat surface-reflectance time series at Monitoring Trends in Burn Severity (MTBS) sampling locations within a large interior-Alaska fire scar, yielding a sample of 90 persistence episodes comprising fully observed and right-censored lifetimes. Weibull parameters are estimated under both maximum likelihood and weighted likelihood, with parametric bootstrap resampling used to obtain confidence intervals. Agreement between model-based and empirical survival behavior is assessed using Kaplan–Meier and Turnbull estimators and a Weibull probability plot. The fitted model indicates a median persistence of approximately 2.7–2.9 years and a 90th-percentile persistence of roughly five to six years, with the Weibull shape parameter consistently exceeding one, indicating an increasing recovery hazard over time. The proposed framework combines interpretability and robustness under censoring and generalizes naturally to other persistence-type environmental processes observed under incomplete monitoring.
Changes in mean precipitation and the frequency and intensity of extreme precipitation represent one of the most consequential aspects of anthropogenic climate change. This study evaluates a set of statistical distributions for modeling trends across the full precipitation spectrum (low, medium, and extreme daily quantiles). Modeling with a single flexible distribution ensures statistical consistency, thereby avoiding the artificial separation and discontinuity inherent in multi-model approaches. We used time as a covariate for dry-day frequency and sea surface temperature for the wet-day distribution parameters. We applied the methodology to a dense network of over 900 meteorological stations in France, offering a wide variety of climatic regimes, allowing for a robust conclusion. We employed a multi-criterion approach to select the best model, in the first step using the Akaike Information Criterion, and then based on their ability to flexibly capture trends across low, medium, and extreme precipitation quantiles. Our findings highlight that three-parameter distributions (generalized gamma and extended generalized Pareto distribution), particularly with evolving shape parameters, are essential for accurately capturing observed differential changes across the full precipitation spectrum, a flexibility that the two-parameter gamma distribution notably lacked. Although AIC generally favored generalized gamma, both generalized gamma and the extended generalized Pareto distribution demonstrated robust performance. This research underscores the critical need for a multi-criterion model identification framework in nonstationary trend analysis to provide reliable insights essential for hydrological and climate impact assessments.
A critical question in climate science is whether climate model simulations are statistically consistent with observations. If simulations and observations are treated as realizations of Vector Autoregressive (VAR) models, then deciding that simulations and observations came from the same process is equivalent to deciding that the parameters of the respective VAR models are equal. This framework has been developed in parts 1–5 of this series of papers, including extensions to account for annual cycles and radiative forcing. However, the associated tests have been derived under the restriction of equal noise covariances. Previous studies have only allowed unequal noise variances in univariate settings. This paper presents a general test of parameter equality that applies to multivariate models, incorporates external forcing, and does not assume equal noise covariances. Monte Carlo experiments indicate that the test statistic is well approximated by a chi-squared distribution for large degrees of freedom, but that this distribution underestimates upper quantiles when the degrees of freedom are small. This bias can be partially compensated by adopting a more stringent significance level (e.g., using a 1 % level to achieve a nominal 5 % Type I error rate). Applying the method to monthly 2 m-temperature from an observational data set and climate model simulations aggregated over five regional domains reveals that most climate models tested differ significantly from the observational data set, both in their transfer coefficients for radiative forcing and in their AR coefficients, indicating differences in the representation of both internal and forced variability.
The nonparametric estimation of the power spectral density of uniformly-spaced data with missing samples is revisited. Classical estimators, such as the standard periodogram and the Lomb-Scargle periodogram, are biased when samples are missing. The classical method to obtain an asymptotically-unbiased estimator is to take the finite Fourier transform of the standard unbiased estimator of the autocorrelation function. However, the latter estimator is not necessarily positive semidefinite, so its finite Fourier transform can yield negative power spectral density values at some frequencies. To avoid this problem, Gao et al.2021)Gao et al.2021)Gao et al.2021)
An understanding of Earth's past climate can help put current and future changes into historical context. Widely used tree ring-based drought atlases generally target the Palmer Drought Severity Index or other metrics of soil moisture and/or drought risk. These indices reflect contemporaneous meteorological conditions, and it is possible to extract information about temperature and precipitation given the existing reconstructions. Here, we present a fully Bayesian inverse method that infers a joint posterior for monthly mean temperature and precipitation given tree ring-based PDSI reconstructions from the North American Drought Atlas. The method is skillful at reconstructing early twentieth century conditions when compared to instrumental measurements from the CRU TS dataset. Moreover, the reconstructions can capture the complex temporal and multivariate covariance structure between monthly regional temperatures and precipitation. By reconstructing regional temperature and precipitation for the last millennium, we identify the driest and wettest years and decades in each region. Our results highlight the unique nature of the 1930s Dust Bowl drought in central Kansas and the late twentieth century pluvial in the North American southwest.
Fine-scale oceanic fronts are ubiquitous and ephemeral physical features that separate contrasting water masses, creating significant heterogeneity in the physical seascape and plankton distributions. Because phytoplankton community composition (PCC) is a key driver of marine ecosystem functioning, understanding the extent to which fine-scale fronts influence PCC is a critical challenge. However, studying PCC across and within fronts is particularly difficult due to data scarcity and high biophysical variability. We developed a tailored statistical model to characterize PCC within an oceanic front we studied in the Mediterranean Sea. We modeled the frontal community as a finite mixture model with three components: two communities of adjacent water masses and a potential front-adapted community. Each component was further considered as a discrete mixture of an unknown number of multivariate Gaussian sub-components. First, we used an Expectation–Maximization algorithm to estimate the Gaussian parameters and determine the optimal number of sub-components based on in situ datasets of the PCC within a frontal zone and its adjacent water masses. Second, a hierarchical Bayesian approach was applied to estimate the weight of all components within the frontal dataset. Our analysis suggests that within the front a new community component, distinct from those in adjacent water masses, accounts for 70 % of the frontal community, indicating that a specific phytoplankton community can emerge in fine-scale oceanic fronts. Despite the limited number of frontal observations, our Bayesian modelling approach provides statistical evidence of the front's influence on phytoplankton community composition, effectively overcoming data scarcity and high variability.
Accurate modelling of extreme precipitation is vital for predicting future risks and informing adaptation strategies. Here, we compare and evaluate six different extreme value statistical models for hourly to 48 h extreme precipitation in southern Germany, with a primary focus on duration-dependent Generalized Extreme Value (dGEV) distributions. To assess model performance, particularly in capturing tail behavior, we utilize the 50-member single model initial-condition large ensemble of the Canadian Regional Climate Model version 5 for the period 1980–2019. The large sample size of 2000 simulated years enables a robust sampling of extreme quantiles. Using a sub-sampling strategy with 30 to 100 years, we compare the efficacy of Bayesian methodology, in particular Bayesian hierarchical models, against frequentist models (L-moments and Maximum Likelihood Estimation – MLE) in representing the tail risk of 100-year return levels based on limited sample sizes. Hierarchical models allow us to give special emphasis on the dimensionality of the GEV shape parameter, a critical factor for tail behavior. Our findings reveal that a shape parameter varying over durations but fixed across space is beneficial for the prediction of the 100-year return level. The resulting Intensity-Duration-Frequency (IDF) curve shows the highest accuracy and smallest confidence intervals proving its robustness. Compared to the standard GEV estimated by L-moments, our proposed model can reduce the relative error of the 100-year return level from 18.1  % to 8.8  % based on a 30-year sample size. Furthermore, our analysis reveals fundamental limitations of the Anderson-Darling test for extreme value model selection, demonstrating its poor correlation with predictive skill for upper quantiles – a critical finding for climate risk applications.
Soil moisture–temperature coupling (SM–T) significantly influences the frequency and amplitude of heat extremes. It describes how variations in soil moisture affect surface air temperature conditions and vice versa. This study aims to determine the spatial extent and duration of SM–Tin situT
We present a novel spatiotemporal clustering algorithm to extract spatiotemporal events based on their intensity. Our algorithm proceeds in two steps: (1) extracting intensity structures that are spatiotemporally consistent and, (2) separating individual events. We apply the algorithm to a novel drought index over the continental United States from 1980–2021 and show that it captures historical drought events over the continental United States and their spatiotemporal extents. 
We present a novel flexible bi-level spatiotemporal clustering algorithm to extract events based on their intensity and spatiotemporal structures. Our algorithm consists of using (i) a novel space-time k-means clustering to obtain spatiotemporally coherent intensity clusters, and (ii) a density-based spatial clustering of applications with noise (DBSCAN) to spatiotemporally section the intensity clusters into individual events. We discuss the development of the algorithm, the selection, tuning and meaning of the parameters within each step, as well as its validation. Finally, we apply the algorithm to a spatiotemporal drought index, standardized vapor pressure deficit drought index (SVDI), over the continental United States (US) from 1980–2021 and show that it captures historical drought events over the continental United States and their spatiotemporal extents.
The aim of this study is to provide a probabilistic gust analysis for the region of Germany that is calibrated with station observations and with an interpolation to unobserved locations. To this end, we develop a spatial Bayesian hierarchical model (BHM) for the post-processing of surface maximum wind gusts from the COSMO-REA6 reanalysis. Our approach uses a non-stationary extreme value distribution for the gust observations at the top level, with parameters that vary according to a linear model using COSMO-REA6 predictor variables. To capture spatial patterns in surface extreme wind gust behavior, the regression coefficients are modeled as 2-dimensional Gaussian random fields with a constant mean and an isotropic covariance function that depends only on the distance between locations. In addition, we include an elevation offset in the distance metric for the covariance function to account for differences in topography. This allows us to include data from mountaintop stations in the training process and to utilize all available information. The training of the BHM is carried out with an independent data set from which the data at the station to be predicted are excluded. We evaluate the spatial prediction performance at the withheld station using Brier score and quantile score, including their decomposition, and compare the performance of our BHM to climatological forecasts and a non-hierarchical, spatially constant baseline model. This is done for 109 weather stations in Germany. Compared to the spatially constant baseline model, the spatial BHM significantly improves the estimation of local gust parameters. It shows up to 5 %
Stochastic weather generators are efficient statistical models producing synthetic weather series by replicating key statistical properties without the computational cost of physical models. However, for applications requiring temperature simulation over a large area, challenges arise due to non-stationarity over time and spatial-temporal dependencies. This paper introduces a daily stochastic weather generator for temperature with arbitrary spatial resolution. The non-stationarity issue is addressed using a decomposition method to separate deterministic terms (trends and seasonality), from the stochastic part representing the underlying climate variability. We extend the existing local decomposition method to extrapolate to any point in space. The spatial-temporal dependence is modeled through a Gaussian field with a non-separable covariance function, accommodating complex interactions between time and space. Our generator, calibrated on a few French weather stations, is validated using several spatio-temporal indicators. First, we evaluate the generator's performance at the fitting stations, comparing simulated and observed indicators. Subsequently, we compare our spatial simulations to a high resolution gridded observation dataset. Results demonstrate that the proposed generator accurately captures the observed spatio-temporal statistics, even for extreme events such as large scale persistent heat waves.
We present a lightweight stochastic weather generator (SWG) based on a multisite hidden Markov model (HMM) trained on a large area with French weather station data. Our model captures spatiotemporal precipitation patterns with a strong emphasis on seasonality and the accurate reproduction of dry and wet spell distributions. The hidden states serve as interpretable large-scale weather regimes, learned directly from the data without requiring exogenous inputs. Compared to existing approaches, it offers a robust balance between interpretability and performance, particularly for extremes. The model architecture enables seamless integration of additional weather variables. Finally, we demonstrate its application to future climate scenarios, highlighting how parameter evolution and extreme event distributions can be analyzed in a changing climate.
Studies that investigate the effects of meteorological fluctuations on varying multi-disciplinary outcomes often depend on analysis of high-frequency sensor data from automatic monitoring stations in different locations. The validation of such spatial time series requires attention given that they are susceptible to multiple forms of error. Existing validation techniques tend to cater to detection of only one form of outlier in isolation, lack robustness, or fail to optimally leverage the strong between-series correlation that often prevails in high-frequency meteorological data exhibiting multiple seasonalities. To address these shortcomings, two adaptations were made to an existing procedure, for more powerful outlier detection in strongly correlated high-frequency time series, using a distributional approach. The modified technique was tested in a simulation study and was also applied to a real univariate spatial set of hourly air temperature series from the South African Air Quality Information System. In both instances, the effectiveness of the technique in detecting outliers was assessed relative to procedures lacking either or both adaptations. The results show the modified procedure to be most comprehensive in the simultaneous detection of multiple forms of error, including solitary spikes, shifts in the series mean, and irregularities in the diurnal pattern. Furthermore, the method is generalizable to any
The Australian Bureau of Meteorology uses linear methods and provides seasonal forecasts expressed as the probability of exceeding median rainfall. I use expanded methods, including more ocean data and deep learning neural networks that provide nonlinear estimates of the rainfall as measured by rain gauges.
Forecasting rainfall into the next season remains highly challenging and is normally presented in terms of probabilities rather than the expected rainfall as measured by rain gauges. I show here that, in favourable cases, for the selected times of the year and selected geographical regions, it is possible to obtain useful quantitative forecasts of rainfall with a series of relatively simple steps. One such instance explored in this work is the prediction of austral springtime rainfall in SE Australia regions predominantly based on the surrounding ocean surface temperatures during the winter.In the first stage, I search for predictors by exploring correlations between the target rainfall and ocean surface temperatures at earlier times. In addition to standard ocean climate indicators such as El Niño or the Indian Ocean Dipole, other typical patterns of variation are captured in terms of the temperatures of selected ocean areas. When characteristic patterns of correlation are discovered, they are included in the predictor selection in the form of expansion in terms of the empirical orthogonal functions (EOFs). EOF expansions can provide very strong signals. For example, in the case of the Indian Ocean, during the winter, the dominant EOF shows a stronger correlation with future rainfall than the commonly used Indian Ocean Dipole.The technical part of the forecast model is provided by deep learning artificial neural networks, where I use the information sources with the strongest correlation in relation to the historical rainfall data as the inputs. The networks are trained on past rainfall data, and the output is a quantitative forecast based on the current state of the predictors. The resulting hindcasts appear to be accurate for September and October and less reliable for November. I also present model forecasts for rainfall during the 2024 austral spring in the selected SE Australia regions.
In the study of the global climate, ocean temperature estimates use sea surface temperature (SST) anomalies instead of marine air temperature (MAT) anomalies. A key question to ask is whether biases result from this choice. In this article we employ hierarchical statistical models to investigate spatiotemporal differences between SST and MAT and their anomalies in the tropical Pacific. The analysis uses observations from the Tropical Atmosphere Ocean (TAO) buoy network and the ERA5 data product. Our spatiotemporal modeling approach accounts for missing data in the observation network and allows for full uncertainty quantification. Our findings indicate evidence that SST and MAT are interchangeable in the tropical Pacific when we calculate seasonally adjusted monthly anomalies.
By the end of 2023, renewable sources covered 63.4 % of the total electric-power demand of Chile, and, in line with the global trend, photovoltaic (PV) power showed the most dynamic increase. Although Chile's Atacama Desert is considered to be the sunniest place on Earth, PV power production, even in this area, can be highly volatile. Successful integration of PV energy into the country's power grid requires accurate short-term PV power forecasts, which can be obtained from predictions of solar irradiance and related weather quantities. Nowadays, in weather forecasting, the state-of-the-art approach is the use of ensemble forecasts based on multiple runs of numerical weather prediction models. However, ensemble forecasts still tend to be uncalibrated or biased, thus requiring some form of post-processing. The present work investigates probabilistic forecasts of solar irradiance for regions III and IV in Chile. For this reason, eight-member short-term ensemble forecasts of solar irradiance for the calendar year 2021 are generated using the Weather Research and Forecasting (WRF) model; these are then calibrated using the benchmark ensemble model output statistics (EMOS) method based on a censored Gaussian law and its machine-learning-based distributional regression network (DRN) counterpart. Furthermore, we also propose a neural-network-based post-processing method, resulting in improved eight-member ensemble predictions. All forecasts are evaluated against station observations for 30 locations in the study area, and the skill of post-processed predictions is compared to the raw WRF ensemble. Our case study confirms that all studied post-processing methods substantially improve both the calibration of probabilistic forecasts and the accuracy of point forecasts. Among the methods tested, the corrected ensemble exhibits the best overall performance. Additionally, the DRN model generally outperforms the corresponding EMOS approach.