Accurate wind speed prediction is essential for renewable energy planning. While machine learning methods often provide strong predictive accuracy for ordinary observations, they frequently underperform in the upper tail of the distribution, particularly for extreme wind events. This study proposes an applied spliced framework integrating Extreme Gradient Boosting (XGBoost) with Peaks-Over-Threshold extreme value modeling to predict the conditional distribution of daily wind speed given that it falls within the retained operational range of 3–25 m/s, rather than the complete daily wind speed distribution. For each station day, an XGBoost regressor estimates the centre of the non-extreme regime, a Platt-scaled XGBoost classifier estimates the conditional probability of exceeding the threshold, and a truncated generalized Pareto distribution (TGPD) models daily exceedance magnitudes between threshold and upper truncation bound. Out of fold residual bootstrap sampling converts the XGBoost bulk prediction into a predictive distribution, and Monte Carlo simulation combines the bulk and tail components on the same daily probability scale. The approach is applied to daily observations from 35 Bangladesh Meteorological Department (BMD) stations during 2000–2020. Relative to an XGBoost only residual bootstrap benchmark, the spliced model reduced the reported continuous ranked probability score (CRPS) from 1.0195 to 0.9926 (2.64
Accurate estimation of extreme rainfall under changing climatic conditions remains a critical challenge for hydrological design and flood risk assessment in urban and coastal regions. This study presents a hybrid bias-correction framework to improve the reliability of extreme rainfall projections derived from CMIP6 climate models for the Kovalam Basin, a flood-prone coastal urban basin in Chennai, India. The approach combines Empirical Quantile Mapping and Quantile Delta Mapping. Extreme rainfall behavior is explicitly modeled using the Generalized Pareto Distribution, enabling robust representation of the upper tail while preserving the projected climate change signal. The framework was evaluated using 26 CMIP6 models over a historical baseline (1955–2014), followed by validation and future projections extending to 2100 under SSP2-4.5 and SSP5-8.5 scenarios. Bias correction improved model performance, with RX1 bias reduced from + 18.9 mm/day to + 5.4 mm/day and better upper-tail representation, reflected by an increase in the 99th percentile from 43.9 mm/day to 51.8 mm/day. A multi-criteria stability screening framework identified four reliable models, namely BCC-CSM2-MR, CanESM5, GFDL-ESM4, and IPSL-CM6A-LR, for ensemble projection. Future projections indicate consistent long-term intensification of extreme rainfall, with ensemble median RX1 increasing by 15.8
A spatio-temporal analysis was performed on wildland fire intensity measurements (thermal temperature data) from a prescribed burn at an ecological research center in Georgia, to better understand the variables and environmental conditions that promote fire spread in a longleaf pine ecosystem (Pinus palustris Mill.). This ecosystem is prevalent in the Southeastern U.S. and prescribed burning is a vital tool for maintaining the ecological function and biodiversity unique to this forest. Measuring fire behavior (spread and intensity) is critical for advancing fire behavior modeling, understanding fire effects and supporting initiatives to support public safety. The infrared cameras used to measure fire intensity had limitations allowing them to only record temperatures strictly above 300 ^∘ C, where plant material combustion mainly occurs. This censoring of temperature values below 300 ^∘ C posed a challenge that initiated this study, which was overcome by augmenting the data set by extrapolating temperatures down to the day’s ambient temperature. This extrapolation is a major focus of this analysis. Spatio-temporal models were applied to these data, and several aspects of the models were evaluated, such as the chosen hyperparameters, the spatial covariance structure, and the prior distributions for the parameters. A sensitivity analysis tested the robustness of the model to several heuristic choices of tuning parameters allowing for multiple reconstructions of the censored data. The model fit to the reconstructed data sets containing extrapolated temperatures appeared to give relevant insight into fire behavior phenomenon. The effects of the covariates in the model, such as wind speed, wind direction, their interaction, and several vegetative fuel variables, such as coarse and fine fuels were significant. These statistical models provide more complete datasets of fire intensity that can be used to evaluate physics-based fire-atmosphere models of fire behavior.
Discriminating among competing statistical models for experimental data is of interest for physical science. While right-skewed experimental measurements commonly approximate a lognormal distribution, among many others, substantial deviations occur at the extremes of the distribution, suggesting that alternative models may provide better fits in certain contexts. In many practical contexts, the choice of a single underlying distribution has significant implications, such as estimating the proportion of dwellings exceeding regulatory reference radon levels. We employ a Bayesian approach to model selection, utilizing Bayes factors and posterior model probabilities to compare suitable distributions applied to radon concentration data from Gran Canaria. Using non-informative priors appropriate for shape-scale family distributions, we calculate marginal densities and derive evidence metrics that transcend the limitations of conventional frequentist goodness-of-fit tests. We found strong evidence against the Gamma model and mild/weak evidence for the LogLogistic over the LogNormal distribution. We also offer a Bayesian model averaging strategy that weights the LogLogistic and LogNormal distributions by their posterior probability to create BMA mixture that improves dwelling exceedance frequency estimates across regulatory reference levels.
This paper explores variance or mean squared error estimation in the design-based inferential framework, focusing on a broad class of estimators and sampling schemes commonly employed in spatial surveys for both continuous populations and finite populations of areas. Specifically, the use of the pseudo-population bootstrap is proposed. It involves constructing a spatial pseudo-population from the sample data using the inverse distance weighting interpolator and then selecting bootstrap samples from the pseudo-population using the same sampling scheme adopted to select the original sample. The properties of the resulting variance or mean squared error estimators are theoretically investigated and empirically assessed through a simulation study. Two case studies are considered.
