Meteorological Applications最新文献

Chongqing is one of the regions in China most frequently and severely affected by potato late blight (PLB), which is a fungal disease caused by phytophthora infestans (PI). To improve PLB occurrence forecast skills in this area, a 1–10 day forecast system (PLBOFS-CQ) based on the CARAH model and the intelligent grid forecast (IGF) of air temperature (AT) and relative humidity (RH) issued by the Chongqing Meteorological Observatory has been developed, showing certain forecast skill. However, IGF errors inevitably exist and increase with forecast lead time, limiting the forecast accuracy. To address this issue, this study investigated an ensemble forecast method for PLB occurrence based on the CARAH model. First, error distribution characteristics of IGF were analyzed, providing a comprehensive understanding of the related forecast uncertainties. On this basis, an error variance–dependent random perturbation method has been developed to generate 200-member IGF ensembles. Long-term verification showed that this perturbation method is reasonable and applicable. Building on this, ensemble mean forecasts (EMF), ensemble quantile forecasts (EQF), and ensemble probability forecasts (EPF) for PI infection have been developed and tested. Among these, maximum EQFs performed best, significantly outperforming the control forecast. The averaged threat score (TS) for infection timing improved by 92.7% at 1–3 day and 34.6% at 4–10 day lead times, whereas improvements for the timing when the Conce score reached 4 after infection were 220.3% and 63.8%, respectively. EPF also demonstrated useful skill, with probabilistic forecasts providing practical references for users. Future work will focus on extending applications and developing an operational ensemble forecast system for PLB occurrence in Chongqing. More broadly, this work demonstrates the potential of ensemble forecast method in agricultural meteorology and provides a pathway for advancing disease forecasting and management in other crop systems.

Caron, J.-F. and B. Casati. 2025. “On the Reliability of Surface Observations and the Pitfalls of Verification Against Own Analyses.” Meteorological Applications 32, no. 6: e70129. https://doi.org/10.1002/met.70129.

The article by Bélair et al. (2003) is cited solely in Section 2.

We apologize for this error.

Accurate forecasting of daily precipitation is critical for agricultural planning and effective water resource management. This study evaluates the capability of machine learning (ML) and deep learning (DL) models to predict daily precipitation using 40 years (1983–2023) of data from five synoptic stations in western Iran. Seven models were tested: Multiple Linear Regression (MLR), Polynomial Regression (PR), Support Vector Regression (SVR), Decision Tree (DT), Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Q-learning with Long Short-Term Memory (DQN-LSTM). Each model was trained on 10-day input sequences to predict precipitation with a one-day lead time, capturing short-term temporal dependencies. Model performance, assessed using R² and RMSE, varied across stations, with DQN-LSTM achieving the best results, explaining over 84% of daily precipitation variability and yielding the lowest RMSE values. Although PR, RF, and XGBoost provided reasonable accuracy, DT and SVR underperformed. However, it is important to note that the models that achieved the best RMSE and R² may not necessarily perform as well in predicting maximum precipitation values at stations. In general, all forecasting methods tend to underestimate the R95p index across stations. Nevertheless, the DQN-LSTM model demonstrates superior overall skill in predicting extreme precipitation indices such as R95p and RX1day. However, for the frequency of extreme precipitation days, the predictions from PR, DT, RF, and XGBoost exhibit closer agreement with the observed values. These findings demonstrate the potential of hybrid DL models like DQN-LSTM to improve both overall forecast accuracy and extreme event prediction, providing valuable insights for water management and disaster mitigation in regions with variable climates such as western Iran.

Two tropical cyclones (TCs)—Phailin and Lehar over the Bay of Bengal (BoB) in 2013 exhibit contrasting rapid intensity changes near landfall, despite forming in similar synoptic environments. Phailin underwent a rapid intensification of ~70 knots between 10 and 11 October 2013, while Lehar rapidly weakened by 30 knots between 27 and 28 November 2013. This study investigates the effects of environmental factors such as vertical wind shear (VWS), the intrusion of cold/dry air, and antecedent land surface conditions (soil moisture and soil temperature; SM/ST) using the cloud-resolving configuration of the Hurricane Weather Research and Forecasting (HWRF) model (at 27/9/and 3-km resolutions). Phailin was characterized by a robust vortex that resisted disruption due to low VWS (10 knots) and modulated its surrounding environment. Whereas Lehar encountered a tilted vortex due to significant VWS (20 knots) and intrusion of mid-level cold, dry air linked to a nearby subtropical high, which weakened its convection/intensity. Cold, dry air alone had a limited impact on storm structure unless accompanied by VWS, which allowed environmental influences to penetrate the core. To quantify the influence of SM/ST, a series of sensitivity experiments were conducted by interchanging them between the two cyclones under similar synoptic backgrounds. Substituting Lehar's land surface conditions into Phailin's simulation showed minimal impact on Phailin's peak intensity, while altering Lehar's surface variables delayed its rapid weakening by 24 h and advanced landfall by 6 h. The study highlights that antecedent land conditions significantly affect storm characteristics even when interacting with land before landfall, highlighting the importance of accurate land surface initialization for intensity forecasts.

Ensembles provide a wealth of information to aid forecasters in their day-to-day operations, but with increasing ensemble size and complexity, there is rarely time to fully interrogate their outputs. Clustering ensemble members into distinct scenarios based on the co-location of hazardous weather features has previously shown promise when applied to global ensemble outputs. However, it is currently unclear whether further value can be gained when applying clustering to convection-permitting ensemble (CPE) outputs. This study compares precipitation clusters between the operational MOGREPS-G driving ensemble and the nested MOGREPS-UK CPE run at the (UK) Met Office during summer 2023. When applied over the UK domain, CPE clustering does not provide clear value compared to global ensemble clustering. Instead, clusters become increasingly similar with leadtime, strongly indicating that CPE clusters are most sensitive to the synoptic forcing common between the two ensembles and that the presence of convective-scale detail has little influence. However, when focussed on a region impacted by hazardous convection, CPE clustering identified distinct precipitation scenarios and provided improved probabilistic value compared to driving-ensemble clustering. Finally, by comparing clusters with radar observations, it is demonstrated that the fraction of members supporting a particular scenario is a reliable quantitative prediction of the probability that the given scenario will be the most accurate. We recommend that global ensemble clustering is sufficient over larger domains, while CPE clustering is most useful when applied at regional scales.

Meteorological station networks are often not homogeneously distributed across geographical verification domains, and usually unpopulated regions (such as deserts or forested regions) are less observed than densely populated regions (such as agricultural regions or cities). Therefore, spatially aggregated verification scores evaluated against station measurements are often dominated by the forecast performance in the regions with a denser observation network. In this study, we explore some solutions used in operational practices for reducing the effects of station network geographical inhomogeneity on spatially aggregated verification scores. The effects of network inhomogeneities on aggregated verification scores is first illustrated over Canada and high latitudes. Thinning the verifying observations to a less dense yet spatially homogeneous network (e.g., considering one station every 1° × 1° latitude–longitude sector) addresses the inhomogeneity issue, but not optimally, since it impoverishes the verification sample. In order to fully exploit the observation network, scores are spatially aggregated by applying a weight to each station, where the weights are inversely proportional to the network density around the station. The weights are evaluated by a Gaussian kernel: we describe a methodology and provide the optimal influence radius, evaluated for the SYNOP station network for different regions around the globe. We conclude that the Gaussian weighting provides more reliable results than thinning, and more representative results than considering the whole (inhomogeneous) station network.

Drought, a significant natural hazard, continues to pose considerable threats to agriculture, particularly in arid and semi-arid regions. Timely and accurate monitoring of drought conditions is essential for effective mitigation and adaptation strategies. This study evaluates the efficacy of three remote-sensing-based drought indices: VCI, TCI, and VHI in detecting and monitoring agricultural drought in the Saurashtra region of Gujarat. The research employs MODIS (moderate resolution imaging spectroradiometer)-derived NDVI (normalized difference vegetation index), and LST (land surface temperature) data to compute the indices. To validate these remotely sensed indices, their values were correlated with the standardized precipitation index (SPI) calculated for 3-, 6-, and 12-month reference periods using precipitation data from the India Meteorological Department (IMD). Furthermore, the spatial distributions and index values were compared between 2002, identified as a drought year by IMD, and 2023, considered a normal reference year. The results indicate that VHI shows the strongest correlation with SPI-6 (r = 0.67), followed by SPI-3 (r = 0.49) and SPI-12 (r = 0.40). This finding aligns with the Standardized Precipitation Index User Guide (WMO-No. 1090, World Meteorological Organization), which recommends using SPI-6 for agricultural drought assessment. Both VCI and TCI exhibit a moderate correlation with SPI-6 (r = 0.62 and 0.56, respectively) but weaker correlations with SPI-12 (r = 0.39 and 0.37). The spatial comparison of VCI, TCI, and VHI between 2002 and 2023 demonstrates that VHI effectively captures the intensity and extent of drought, as it integrates vegetation and thermal stress. Overall, the study highlights the potential of VHI as a reliable, remote-sensing-based drought indicator that provides timely information on drought severity and spatial extent, particularly in arid and semi-arid regions. Integrating VHI with soil-moisture data could yield an even more robust composite drought index for policymakers and agricultural stakeholders to support strategies that mitigate the adverse impacts of drought on crop production and livelihoods.

The Indian subcontinent shows significant spatial and temporal variability of precipitation. A small change in precipitation frequency and its distribution may affect agriculture and water resources and can lead to extreme events such as floods and droughts. In the present study changing precipitation characteristics over different meteorological Indian sub-regions are presented. Indian Meteorological Department (IMD) gridded precipitation and ECMWF Reanalysis 5th Generation (ERA5) reanalysis data during 1970–2020 are considered. Furthermore, the Theil–Sen slope test and Pettit's test are used for calculating the magnitude of trend and change point respectively for the number of precipitating days and associated precipitation over India and its sub-regions. Early arrival of the wettest day (day with maximum precipitation) is observed over northeast India and northern central northeast India, while the increase in the duration of the rainy season over northwest India is observed. Extension of higher precipitation to July–August–September–October is distinct over India except for the central northeast. Change point detection shows these changes occurred mostly after 1996. The decreasing precipitation trend across northeast and central northeast, while the increasing trend over northwest India reflects a westward strengthening of the monsoon precipitation. Additionally, greater moisture transport from the Arabian Sea and Bay of Bengal is detected in the recent period (1997–2020), which may be the reason for higher precipitation over northwest India. Overall, the results will aid in understanding how climate change affects the Indian summer monsoon, which will support policy making and adapting water management techniques.

The emergence of high-resolution numerical weather prediction (NWP) systems over recent decades has brought new verification challenges, namely accounting for the “double penalty” effect. While spatial verification methods have been developed to mitigate this issue, they generally provide domain-wide performance assessments, potentially obscuring spatial heterogeneity in the NWP performances. This study introduces a novel methodology for evaluating the NWP performances at the local scale within a neighborhood-based framework. Local contingency tables are constructed for each cell of the grid, populated with events occurring within a defined neighborhood window, allowing for the compensation of spatial location errors. These local contingency tables are then temporally aggregated across a set of forecasts to produce a temporal local contingency table at each grid point, thereby enabling localized performance assessment. The methodology was applied to a large region centered in Southern Quebec using forecasts from six NWP systems (GDPS, RDPS, HRDPS, GFS, NAM, and RAP) over a 2-year period (2022–2023). Analyses were conducted across four precipitation intensity thresholds (0.1, 5, 10, and 25 mm/6 h) and three forecast lead-time categories (Days 1–2, 3–4, and 5–7 combined, depending on data availability). Four metrics were employed in the evaluation: Bias, false alarm ratio (FAR), probability of detection (POD), and equitable threat score (ETS). The performance is primarily governed by the precipitation intensity threshold, with forecast skill deteriorating as the threshold increases, particularly, for intense and extreme events. Although forecast lead-time has a secondary yet nonnegligible influence, spatial variability of metric values becomes increasingly pronounced at higher intensity thresholds, despite certain limitations in evaluating extreme precipitation events. Notably, the evaluation at the local scale and the delineation of homogeneous regions proved particularly valuable at the 5 mm/6 h threshold, underscoring the relevance of localized verification approaches for moderate precipitation events.

This study proposes a pattern-referencing model for hourly temperature forecasting in coastal regions, specifically designed for scenarios with missing data. The Chiayi–Tainan coastal plain in Taiwan exhibits pronounced spatiotemporal temperature variations driven by sea–land breezes, topography, and solar radiation, impacting real-time decision-making in industries such as aquaculture, agriculture, and tourism. The proposed model directly utilizes all available input data without requiring prior imputation or specialized pretraining. In a multistation study involving 14 weather stations, the model employs a weighted K-nearest neighbors (WKNN) approach, using a masked Euclidean distance and the Dudani weighting scheme. The optimal configuration (look-back length = 1, number of neighbors = 18) achieved mean absolute errors of 0.35°C–0.59°C and root-mean-square errors of 0.45°C–0.86°C across diverse weather scenarios, outperforming both persistence forecasts and an autoregressive integrated moving average (ARIMA) model. The model performs best under low-temperature conditions but shows a slight tendency to underestimate at high temperatures; nighttime forecasts are the most stable, while daytime errors are larger. Even with missing station data, the model maintains its predictive capability, offering decision-makers more reliable hourly forecasts in resource-limited networks with unstable data availability, and enabling policymakers to build early-warning systems that help coastal communities and industries respond to extreme temperature events.