Quantifying rainfall seasonality requires a representation that respects the cyclic structure of the annual calendar while allowing estimation, robustness assessment, and inference under temporal dependence. We develop a divergence-based framework in which rainfall timing is represented as a circular probability distribution and seasonality is measured by a phase-invariant convex divergence from circular uniformity. The functional is estimated by circular kernel density methods, its local sensitivity is characterized through influence functions, and inference is calibrated using a circular stationary bootstrap with a data-driven reference block length and explicit sensitivity analysis. Simulation experiments examine concentration, multimodality, phase rotation, temporal aggregation, localized extremes, and serial dependence. RSI responds to overall departure from uniform timing rather than to the number of peaks alone, remains numerically invariant under circular phase shifts, and retains information in symmetric multimodal settings where the mean resultant length can be attenuated by directional cancellation. Localized-extreme experiments produced context-dependent responses consistent with the bounded-influence motivation, but did not indicate uniform robustness superiority. Resolution-controlled simulations indicate that RSI retains about 94
Fine-resolution land surface temperature (LST) is essential for irrigation scheduling, urban heat mapping, and heat-health risk assessment. MODIS provides daily LST, but its 1 km pixels smooth thermal differences among surface types; Landsat resolves these patterns at 30 m but revisits every 16 days. Deep-learning downscaling metrics can be misleading because a network may reduce RMSE by reproducing the resampled MODIS field without recovering fine-scale detail. SCTV-ResNet addresses this by design. A coarse calibration head corrects systematic MODIS–Landsat bias, while a residual head adds fine-scale detail from vegetation and terrain predictors. The same architecture was evaluated under three prespecified holdouts (an unseen year, unseen locations, and their combination) using 787 MODIS–Landsat scene pairs over Ranchi and Khunti, Jharkhand, India, acquired during 2021–2026. Test RMSEs were 1.70, 1.37, and 1.81 °C, representing ∼60–71
As the climate warms, understanding patterns of heat risk and related impacts becomes essential for informed decision-making and adaptive planning, particularly in hot-arid environments where climate regimes are characterized by frequent occurrence of multiple heat types. The impacts of heat hazards not only vary among heat types (e.g., hot days, hot nights, and heat waves) but also within each type. Building on this and using data from Saudi Arabia, a multidimensional heat-hazard index was proposed to quantify the compound characteristics within each heat-hazard type. Results showed that the frequency, intensity, and duration metrics capture distinct characteristics of heat events and relying on a single metric does not fully represent the associated risks. The developed multi-dimensional index demonstrated adequate performance in accounting for multiple heat-hazard dimensions and further highlighting spatial differences. Thermal risks for each heat hazard were examined using a combination of risk-triangle analysis, bivariate mapping, and ternary diagrams. This integrated framework enabled a more in-depth analysis of interactions among the three risk dimensions—hazard, exposure, and vulnerability—and provided a valuable tool for informed decision-making. Hot nights posed higher risks than other hot extremes, and vulnerability–hazard interactions tended to play a greater role in shaping overall risks across hazard types, with varying contributions. Overall, the findings suggest that tailored approaches—both hazard-specific and location-specific—are essential for effective risk reduction interventions. The applied framework contributes to the first systematic profiling of heat risks in Saudi Arabia while remaining transferable to environmental and integrated risk-impact assessments in other hot-arid regions.
In the context of multi-objective water resource allocation and forecasting future water demand in irrigation districts, it is critical to understand the evolution of the water-use structure and its influencing factors, particularly in these water-scarce areas. This study presents an integrated analysis framework encompassing “temporal evolution-structural equilibrium-spatial agglomeration” based on various methods, such as information entropy, the Lorenz curve, the Gini coefficient, Moran’s Index, Getis-Ord Gi* hot-spot analysis, and logarithmic mean divisia index (LMDI) decomposition, to investigate the changes and driving factors of water use patterns in Yuncheng City, China, during 2007–2023. Key findings include: (1) Total water consumption demonstrates a fall-rise trajectory, peaking in 2016, with agricultural water use consistently accounting for 75–85
Coastal aquifers are increasingly vulnerable to seawater intrusion, climate change, and intensive groundwater extraction, while available monitoring data are often limited in both quality and quantity. These constraints restrict the application of physics-based models and complicate reliable risk assessment and management.This study develops and validates a data-driven framework for short-term forecasting of piezometric level and salinity dynamics in coastal aquifers, with explicit consideration of model uncertainty. Five regression and machine-learning methods were applied to datasets from two Mediterranean coastal aquifers: the Señorío aquifer in Marbella, southern Spain, where electrical conductivity and piezometric level were modeled, and the Sierra de Mijas aquifer in Torremolinos, eastern Spain, where piezometric level was modeled. Forecasts were produced for 3-, 6-, and 12-month horizons using limited historical records, and model performance was evaluated by metrics capturing accuracy, trend reproduction, and probabilistic characteristics. Results show that the ARDL model performs robustly for groundwater time series with small oscillations, whereas non-linear approaches, including recurrent neural networks, are more effective in capturing oscillatory behavior and short-term fluctuation. The Gaussian Process framework yields reliable prediction intervals and introduces a normalized measure of relative model uncertainty that supports model comparison and risk-informed interpretation of forecasts. Despite limited training data, all models achieved effective one-year forecasting performance. The inferred correlations between piezometric level, electrical conductivity, rainfall, recharge, and extraction are consistent with established physical understanding of coastal aquifer systems. Overall, the proposed framework provides a practical tool for supporting groundwater risk assessment and adaptive management in data-scarce coastal aquifers, with potential applications in early-warning systems and operational decision-making.
Structural uncertainty arising from constitutive models critically affects the accuracy of dense non-aqueous phase liquid (DNAPL) migration simulations in low-permeability media. Traditional multi-model approaches, such as Bayesian model averaging (BMA), suffer from prior limitations and are difficult to apply under complex field conditions. This study introduces Bayesian stacking (BS) for the structural uncertainty analysis of DNAPL migration models in low-permeability media. First, Markov chain Monte Carlo (MCMC) simulation is employed to quantify parameter uncertainty. Subsequently, BS is applied to analyze structural uncertainty and generate weighted ensemble predictions for four constitutive model combinations, with results compared with the conventional BMA. The research is based on two cases of DNAPL migration in a 2D laboratory sandbox experiment and an idealized 3D heterogeneous model, respectively. The results show that the model weight distribution of BS based on leave-one-out (LOO) cross-validation and logarithmic score optimization is more reasonable. It can effectively alleviate the problem of equifinality in BMA. BS has broad application prospects in groundwater pollution simulation and risk assessment.
Tide gauges are critical components of coastal monitoring infrastructure, yet their reliability under severe environmental conditions remains insufficiently understood. This study investigates the drivers and temporal dynamics of recurrent failures in two long-term tide-gauge installations using survival analysis and stochastic process modelling. Recurrent failure hazard was analysed using stratified Cox proportional hazards models that incorporated wind speed, wave height, tidal range, and precipitation. A significant negative wind–wave interaction indicated that the association between wind speed and failure hazard weakened as wave height increased, rather than representing a uniformly additive storm effect. Threshold-based exceedance models did not outperform continuous wind formulations, indicating that continuous representations of wind exposure better described failure risk than the tested threshold formulations. Complementary Weibull accelerated failure-time and piecewise-exponential models suggested a non-constant baseline hazard, with higher hazard rates early in operational spells. The temporal frailty analysis provided only weak exploratory evidence of residual heterogeneity across broad multi-year periods. Finally, station-level failure dynamics were analysed using a Weibull non-homogeneous Poisson process (Crow–AMSAA). Station-level Crow–AMSAA estimates were β = 0.879 at Leixões and β = 1.078 at Sines; the 95 β = 1 at both stations, providing no clear evidence of long-term reliability growth or deterioration. These results emphasise the importance of combined wind–wave forcing in shaping recurrent failure risk and reveal that, although point estimates differed between stations, no statistically supported long-term trend in failure intensity was identified. Together, the proposed modelling framework provides a probabilistic basis for understanding failure processes in coastal monitoring systems.
The elucidation of spatial gradient patterns in rural-urban ecosystem health (EH) and environmental factors (EFs), as well as the intricate multi-gradient driving mechanisms, is essential for the effective management of rural-urban environments. This study employed an improved Vigor, Organization, Resilience, and Services (VORS) model to evaluate the gradient distribution of EH levels within the Urban Agglomeration of the Middle Reaches of the Yangtze River (UAMRYR). Utilizing grey relational analysis and Geographically Weighted Regression (GWR), this study quantified the spatiotemporal co-evolutionary associations between EH changes and EFs across various gradients. The Optimal Parameters-based Geographical Detectors (OPGD) and Generalized Additive Model (GAM) were employed to characterize the gradient-stratified statistical explanatory power of EH sub-dimensions over EFs, and to delineate the nonlinear associative responses of EFs to EH variation. The findings reveal several key insights: (1) From rural areas to urban cores, the spatial correlation among EFs shifts from predominantly low-low clustering to high-high clustering. (2) The spatial patterns of EH and its four sub-dimensions remain relatively stable. Initially, EH levels rise with increasing rural-urban gradients (T1-T9), fluctuating twice before declining overall. (3) The correlation between the rates of change in EH and EFs shows spatial heterogeneity across multiple gradients, with changes in the direction (synergy/trade-off) and intensity of their interactions over time. (4) At various stages and gradients, the principal EH driver of each EF varies across stages and gradients. With increasing rural-urban gradients (T1-T9), EH’s the gradient-stratified statistical explanatory power for LST and PM2.5 generally increases initially and then decreases, whereas for CO2, it steadily declines and stabilizes at a lower level. Nonlinear responses to changes in Ecosystem Health Index (EHI) across rural-urban gradients reveal that the statistical capacity of EH to predict EF variation exhibits spatially differentiated and threshold-governed patterns, with varying environmental mitigation efficiencies across gradient tiers. By elucidating these multi-gradient, non-linear, and spatially explicit dynamics, this study provides a robust statistical-analytical framework that moves beyond one-size-fits-all policies, enabling the design of targeted, place-based strategies for effective rural-urban ecological governance.
Infrastructure siting for hybrid renewable energy systems is inherently a stochastic, multi-criteria problem shaped by uncertain environmental conditions, socio-economic constraints, and equity considerations. Deterministic geographic information system (GIS) approaches often fail to quantify the probabilistic risk of siting errors, limiting their value for decision-makers in dynamic contexts. This study introduces a Stochastic Risk-Calibrated Geospatial Siting framework that integrates conformal prediction-based uncertainty quantification with a hybrid NSGA-II/Bayesian Optimization solver to identify Pareto-optimal sites across techno-economic, environmental, and social-equity objectives. The framework generates calibrated suitability intervals [ ŷ L(x), ŷ U(x)] that capture spatial prediction uncertainty, enabling explicit risk-informed decision thresholds for energy access planing. Three contrasting regions—Cameroon, Canada, and Saudi Arabia—were analyzed under ± 20
This study developed a sequential empirical framework for investigating the dynamic relationship between economic activity and air quality dynamics by integrating autoregressive (AR), vector AR (VAR), and Markov regime-switching models. Industrial production was incorporated as a leading indicator of economic activity to examine whether business-cycle fluctuations are associated with regime-dependent air quality dynamics. VAR and Granger causality tests identified industrial production as a significant leading indicator of the cumulative sub-air-quality index in Taiwan, with a lead time of approximately 4 months. This predictive relationship provided the empirical basis for incorporating industrial production into the proposed sequential framework. The preferred Markov regime-switching specification identified distinct high- and low-pollution regimes and indicated that the associations of industrial production differed across pollution regimes. The proposed framework provides a systematic approach for analyzing the interaction between economic activity and pollution-state-dependent air quality dynamics and offers a scientific basis for future research on environmental risk management and quantitative environmental finance. The framework is readily applicable to industrialized economies where economic activity exhibits predictive relationships with air quality dynamics.
Fine particulate matter (PM2.5) forecasting supports public-health advisories and operational early-warning systems. We present a European, multi-country benchmark for direct, multi-horizon PM2.5 forecasting (1/3/6/12/24 h) that compares statistical, tabular machine learning, and sequence deep learning models under a single, reproducible experimental design. We construct a harmonized hourly dataset (2018–2024) by joining the European Environment Agency (EEA) station measurements with meteorology and station metadata and evaluate two complementary protocols: Protocol A—maximum tabular coverage—and Protocol B—a common sequence-eligible subset enabling cross-paradigm fairness. Across horizons, boosted trees (LightGBM/XGBoost) are consistently strong under Protocol A, while under Protocol B residual long short-term memory (LSTM) variants (with attention at h = 1) are competitive at short horizons and boosted trees dominate at medium–long horizons. Stratified analyses reveal substantial heterogeneity by country and station area, motivating horizon-specific models and stratified monitoring in deployment. We further quantify the coverage–comparability trade-off, report skill vs persistence, and paired significance tests, and provide feature-importance summaries to aid interpretation. All code, configurations, and masks are released for full reproducibility, establishing a transparent baseline for future methodological advances.
Landslide susceptibility assessment (LSA) is an important tool for landslide disaster risk management. However, traditional point-based models have limitations in handling the spatial dependencies of landslide conditioning factors. This study employed a deep learning method of 3-dimensional convolutional neural network (3DCNN). By directly processing landslide conditioning factor data in image form, the model effectively captures the environmental characteristics of the surrounding area. Taking Huizhou City in Guangdong Province as a case study, the 3DCNN model outperformed the 1-dimensional convolutional neural network (1DCNN) and traditional machine learning models (e.g., SVM and RF) in terms of prediction accuracy, precision, F1 score, and AUC, with an AUC value of 0.93. The model’s predictions indicate that areas with low mountainous terrain, sparse vegetation, developed rock fractures, and frequent human activity are high-risk zones, closely aligning with the distribution of historical landslides. Notably, the 3DCNN model identifies that “very high” susceptibility areas cover only 9.78
We present a statistical assessment of out-of-sample correctability limits under observational uncertainty, using GPM IMERG Final V07 sub-daily areal precipitation extremes over the Comunitat Valenciana (eastern Spain) and the dense AVAMET network (556 gauges, 2019–2025). Rather than treating gauges as point truth, we construct a gauge-derived multi-station operational areal reference proxy at 30 min over 126 IMERG cells and build a closed benchmark at 1, 3, and 6 h under year-holdout, province-holdout, and event-fold validation. The benchmark compares raw IMERG (M0), two simple statistical corrections (M1–M2), a continuous LightGBM corrector (M3), a two-stage tail-oriented LightGBM model (M3b), and a conservative patch-based CNN specialist (M4). We further evaluate reference sensitivity under alternative proxy definitions and pseudo leave-one-gauge-out perturbations, and we treat direct probabilistic exceedance modelling as a core benchmark component alongside deterministic correction. Under the main operational-median proxy, M3 is the best global continuous corrector at all scales, reducing event-fold RMSE from 0.461 to 0.360 at 1 h, from 0.801 to 0.578 at 3 h, and from 1.445 to 1.259 at 6 h. However, these gains do not translate into clean recovery of severe and extreme events under fixed operational thresholds. Tail-oriented models improve severe-event skill relative to M3 at 3 h and 6 h, but only modestly, while fixed extreme recovery remains weak across reference perturbations and hard holdouts. The conservative local deep-learning specialist does not materially improve the tail-oriented tabular baseline. About half of the severe and extreme cases occur in cells with the minimum observational support of two gauges, and local jackknife diagnostics show increasing proxy sensitivity in the upper tail. Peak and alignment-oracle diagnostics indicate that residual error is only partly explained by local spatiotemporal misalignment; substantial amplitude underestimation persists even after temporal and neighborhood tolerance, and IMERG often enters far below the operational threshold in observed extreme cases. Importantly, direct probabilistic exceedance modelling emerges as the main positive result: it provides robust out-of-sample risk ranking for severe and extreme exceedance across split families, although operational precision remains limited by event rarity. Overall, the results support partial out-of-sample correctability of IMERG sub-daily areal extremes under an operational areal proxy, while probabilistic risk ranking emerges as the central operational output when clean deterministic recovery of the fixed extreme tail remains unattained. For operational use, the most cautious interpretation is to pair continuous correction with direct probabilistic risk ranking rather than rely on corrected satellite estimates alone for threshold-based event detection in early-warning or hazard-screening applications.
Historically, mountain activities have been associated to accidents, injuries and fatalities. The spatial and temporal context, however, has been largely neglected, with most studies focusing on proximal causes of accidents. The objective of the present study is to analyse the effects of spatial and temporal covariables on the distribution of mountain accidents in a specific area over a span of 11 years. The current dataset includes 572 rescues on Montserrat Natural Protected Area between 2011 and 2021 and comprises 249 climbing rescues and 310 hiking rescues. We assume that mountain accidents follow a Log-Gaussian Cox Process (LGCP) and we consider an empirical analysis of the first-order characteristics. Then we propose a model in which the conditional intensity of the point process depends on some specific spatial and temporal covariables affecting the distribution of this space-time point pattern. We use the inhomogeneous spatio-temporal K-function to estimate second-order properties. Finally, we model the residual spatio-temporal variation as a stochastic process using a space-time covariance function under a separable space-time structure, and we conduct a risk analysis based on the resulting full LGCP through the Value-at-Risk. The results indicate that rescues are clustered over short distances, typically below 100–300 m. Spatial and temporal predictors differ across activities, while risk remains consistently concentrated in specific areas throughout the study period. Overall, the full LGCP model shows a good fit and is able to generate simulations consistent with the observed data.
Random fields provide a flexible framework for modelling spatial variability in environmental, mineral, and geological processes. In mineral exploration data, spatial continuity is often anisotropic because of formation processes such as sedimentary stratification. This type of structure can be represented through geometric anisotropy. Anisotropy parameters are commonly inferred from directional variograms. This study examines the applicability of the anisotropic Hybrid Spectral Ornstein–Uhlenbeck (HSOU) covariance model for multivariate geostatistical modelling under geometric anisotropy. The methodology is applied using collocated zinc (Zn) and lead (Pb) soil samples to support sustainable mining development. Gaussian anamorphosis based on Kernel Cumulative Distribution Estimation (KCDE) was applied to transform the data to an approximately standard Gaussian distribution. Directional empirical direct variograms and cross-variograms were used to characterise anisotropy in the spatial continuity structure. The estimated anisotropy ratios were R_Pb=1.23 , R_Zn=2.53 , and R_Pb-Zn=1.76 . Cokriging predictions obtained using isotropic and anisotropic HSOU models yielded comparable cross-validation performance. Uncertainty in the variogram parameters and predictions was evaluated through Monte Carlo simulations. The substantial uncertainty in the variogram parameters propagated to the cokriging predictions, resulting in considerable uncertainty in the estimates. However, the results indicate that the anisotropic HSOU model produces admissible covariance matrices and stable multivariate predictions under directional heterogeneity, supporting its use in mining geostatistical applications.
The sustainable management of renewable natural resources requires decision-making frameworks capable of accounting for both environmental and economic uncertainty. This paper develops a Real Options framework for fishing activity in which fish populations follow a stochastic Gompertz process and prices evolve according to a mean-reverting Geometric Ornstein–Uhlenbeck process. The valuation problem is formulated as a Hamilton–Jacobi–Bellman equation, solved numerically using a Crank–Nicolson finite difference scheme, and compared with an expected net present value approach under the same economic, biological, and adaptive harvesting assumptions, thereby isolating the effect of the valuation methodology. The results show that both approaches generate similar harvesting policies and population trajectories, with neither suggesting overexploitation over the simulated horizon. Nevertheless, the Real Options framework yields higher project values, primarily due to its treatment of systematic price risk. By linking risk compensation to a spanning asset rather than to an exogenous discount rate, it provides a more internally consistent valuation methodology. The comparison also reveals a trade-off between economic value and biological resilience, as the higher values obtained under the Real Options framework are associated with lower average population levels and greater dispersion in population outcomes. These findings highlight the relevance of Real Options methods for the economic valuation and sustainable management of renewable natural resources under uncertainty.
Short-term wind speed forecasting remains challenging because atmospheric wind fields exhibit strong non-stationarity and stochastic variability, which limit the effectiveness of conventional deterministic forecasting models, particularly under complex or data-constrained conditions. This study proposes a Diffusion-Encoder framework that reformulates forecasting as a conditional diffusion-based generative process and serves as a model-agnostic enhancement layer for existing forecasting architectures. The framework was integrated with five representative deep learning architectures and evaluated using four datasets spanning mountainous and offshore environments under both warm- and cold-season conditions. Performance was assessed across 1-, 3-, 6-, and 9-step forecasting horizons and under progressively reduced training data availability (100
Environmental contamination, particularly in groundwater, has significantly increased in urban areas affected by multi-industrial and commercial operations, as well as agricultural practices carried out in peri-urban areas of Faisalabad. To assess the quality of groundwater in Faisalabad, composite samples (n = 24) were collected at depths of 10–120 feet from rural, urban, industrial, and urban complex zones. The samples were analyzed using an atomic absorption spectrophotometer to assess the content of As, Cr, Cu, Mn, Ni, Pb, and Zn. The source apportionment model identified agricultural practices (29
Flexible and parsimonious parametric models for time-to-event data are valuable when hazards depart from simple monotonic shapes. This article aiming at the development of Flexible Exponential-Type Family (FETF), a probability-scale generator that transforms a baseline cumulative distribution function via an exponential-type mapping, and develop the Exponential-Type Weibull (ETW). To validate the theoretical performance of ETW, we derived a closed-form density and hazard expressions, identifiability results, regularity conditions for likelihood inference, and series representations of moments and mean residual life. To evaluate practical performance, we present a comprehensive Monte Carlo benchmarking design that contrasts ETW with Weibull, Inverse-Weibull, Transmuted Inverse-Weibull and Alpha-Power Weibull families across canonical hazard regimes (increasing, decreasing, unimodal) and classical bathtub scenarios (mixture/shifted generators). Simulation metrics include bias, mean squared error, empirical coverage, convergence rates and model-selection (AIC/BIC) frequencies. The empirical caparison of the proposed method on an anonymized Mexican COVID-19 mortality cohort (n = 106), reporting MLEs, information criteria and robust graphical diagnostics (QQ/PP with bootstrap bands, Cox–Snell residuals, total time on test). Numerical finding show ETW is a compact, viable choice when the true hazard is unimodal or otherwise non-monotonic and when an initial zero hazard at time-zero is acceptable, when initial preeminent hazard is existed, transformed or mixture models are superior. We conclude with practical guidance for applied use, explicit limitations of the illustrative dataset, and concrete extension paths.
Flood monitoring and water level estimation are essential for effective flood management but remain challenging due to sparse station distribution and complex nonlinear dynamics. This study proposes a Sentinel-1 SAR-based machine learning framework that integrates Sentinel-1 observations with a water-level-conditioned Long Short-Term Memory (LSTM) modeling for water level estimation in the Yeongsan River Basin, South Korea. Water body detection used Otsu thresholding and K-means clustering and validated against ESA WorldCover maps. Both methods achieved high accuracy (0.88–0.90) with low false alarm rates (0.02–0.04). K-means outperformed Otsu across most sites, while Otsu showed better performance at tributary stations with small water bodies due to reduced class imbalance. However, both methods exhibited reduced accuracy along land–water boundaries where mixed scattering was prevalent. For water level estimation, Sentinel-1 backscatter (σ⁰VV, σ⁰VH), incidence angle, and day of year were used as inputs to the LSTM model. Model performance exhibited pronounced spatial heterogeneity: tributary sites governed primarily by rainfall-driven hydrological conditions achieved superior performance (R up to 0.89; RMSE < 0.20 m), whereas mainstream sites influenced by dams and weirs showed limited accuracy (R < 0.30). To address hydrologic nonstationary and data imbalance, a water-level-conditioned training strategy was introduced by splitting datasets into high- and low-water level conditions. This approach enhanced model performance at several mainstream sites, although underestimation of peak levels remained due to data imbalance and the relatively low temporal resolution of SAR observations. Overall, the results demonstrate the potential of integrating SAR observations with machine learning for water body detection and water level estimation. The proposed framework provides a practical and data-efficient approach for hydrological monitoring and flood risk management, particularly in data-sparse and hydrologically complex river systems.
