Search Results (674)

This controlled proof-of-concept study evaluated whether interpretable RGB histogram descriptors extracted from smartphone images can be used to estimate nominal syrup addition within a single experimentally prepared honey-adulterant series. Eleven concentration-specific physical mixtures containing 0–50% syrup in 5% increments were prepared, and each mixture was represented by ten technical replicate images acquired under fixed white-light and camera settings while the sample vessel was rotated. Whole-image descriptors and a secondary patch-based representation were analyzed using partial least squares regression (PLSR), support vector regression (SVR), random forest (RF), gradient boosting (GB), and extreme gradient boosting (XGBoost). Nested leave-one-concentration-out validation kept all technical replicate images originating from each concentration-specific preparation within the same fold, thereby preventing replicate leakage between training and test data. Performance metrics were calculated from the eleven outer concentration-level predictions. The primary analysis allowed fold-specific selection from the complete pool of 102 RGB descriptors; PLSR performed best (R²_nested-LOCO= 0.980, RMSE_nested-LOCO = 2.22 percentage points, and MAE_nested-LOCO = 1.96 percentage points), followed by SVR (RMSE_nested-LOCO = 2.74 percentage points). Secondary restricted-feature analyses separately evaluated the G-channel descriptor family, the three mean RGB intensities, and G_mean alone. A performance-weighted cross-model SHAP analysis showed that location and percentile descriptors accounted for 67.7% of the total consensus importance score, with G_p95 ranked first. The G-channel descriptor family consistently outperformed the three RGB means, while G_mean provided a strong but less informative reference. Whole-image descriptors outperformed patch-based representation for four of five algorithms in the matched comparison. These results demonstrate the feasibility of quantitative syrup-addition estimation within the investigated controlled series rather than general honey-adulteration detection. Future studies should include independently prepared samples, broader honey–adulterant combinations, and evaluation of transferability across imaging devices and acquisition sessions. Full article

Economically motivated adulteration of olive oil, coffee and fruit juice is a persistent food-fraud problem for which Fourier-transform infrared (FTIR) spectroscopy with chemometrics offers rapid screening. Linear partial least squares (PLS) is interpretable but cannot capture non-linear mixing; neural networks add flexibility at the cost of becoming black boxes. We evaluated a Kolmogorov–Arnold network (KAN), which places learnable univariate functions on its edges and is therefore intrinsically interpretable, against PLS, support-vector regression, random forests, a multilayer perceptron and a one-dimensional convolutional network on three attenuated total reflectance (ATR)–FTIR datasets (olive oil + sunflower oil, coffee + malt flour, orange juice + apple juice; approximately 350, 400 and 400 spectra). All models were compared under identical, leakage-free validation that splits spectra by physical sample. The compact KAN was consistently competitive (cross-validated coefficients of determination (R²) = 0.86, 0.93 and 0.69) and yielded closed-form equations whose variables map to recognised vibrational bands and whose importance ranking agrees with SHapley Additive exPlanations (SHAP; Spearman ρ = 0.86–0.90); symbolic conversion costs no accuracy. We also report the following limits: PLS was strongest where the chemistry was linear (coffee) and the multilayer perceptron was strongest on fruit juice, whose equation is the weakest (R² = 0.47–0.75 across seeds); a parameter-matched perceptron matched the KAN’s accuracy; and leave-one-brand-out validation degraded every model. The KAN is therefore a promising, compact and genuinely transparent alternative under controlled multi-matrix conditions, not a deployment-ready method. Full article

Reliable modelling of monthly air temperature is relevant to station-scale climate assessment and the evaluation of meteorological data-driven models. However, station-scale monthly meteorological datasets often contain correlated and partially redundant predictors because thermal, moisture, precipitation, wind, and seasonal variables are jointly controlled by atmospheric and seasonal forcing. This study conducts an integrated comparative analysis of established regression and machine learning models for monthly mean air temperature modelling in Zonguldak, a humid coastal province in the Western Black Sea Region of Türkiye. Monthly meteorological observations from 2000 to 2022 were used to evaluate eight primary regression and machine-learning models: Partial Least Squares regression, Ridge, Lasso, ElasticNet, Support Vector Regression, Random Forest, Gradient Boosting, and Extreme Gradient Boosting. Ordinary Least Squares (OLS) and Huber regression were additionally included as reference models. The analysis retained the original meteorological predictors and jointly evaluated predictive accuracy, model stability, ablation sensitivity, and model-specific predictor relevance. Reduced-predictor and seasonality-only scenarios were examined to distinguish direct thermal reconstruction from broader climatological predictability. Model performance was assessed using repeated nested cross-validation, bootstrap summaries of performance variability, supplementary rolling-origin validation, and Wilcoxon signed-rank tests with Holm correction. Although the full-predictor models achieved high predictive accuracy, this performance largely reflected the direct thermal information contained in minimum and maximum air temperature. When these thermal predictors were excluded, MAE increased to approximately 1.13–1.22 °C and $R^2$ decreased to approximately 0.93–0.94. The seasonality-only scenario yielded MAE values of approximately 1.27–1.34 °C and R² values of approximately 0.92, indicating that the annual cycle accounted for a substantial proportion of monthly temperature predictability. The additional non-thermal meteorological predictors provided only limited improvement beyond the strong seasonal baseline. Overall, model performance depended on the predictor information available, and no single model family showed a consistent advantage across the evaluated scenarios. These findings highlight the importance of considering predictive accuracy together with model stability and predictor dependence in data-limited station-scale temperature modelling. Full article

This study investigated how detection distance affects spectral models for soluble solids content (SSC) and firmness evaluation in Korla fragrant pears and provides a reference for calibrating standardized indoor non-destructive detection equipment. Two hundred visually intact fruit samples at the early-ripening stage were collected from the Korla production area in Xinjiang. An FS-640 multispectral camera system equipped with a VS-SWR fixed-focus industrial lens (16 mm focal length, F1.8 maximum aperture, 1/2-inch sensor format) was used to acquire fruit reflectance spectra at seven vertical lens-to-fruit-surface distances of 90, 100, 110, 120, 130, 140, and 150 cm. A 625-pixel region of interest (ROI) was selected using ENVI at an undamaged equatorial or near-equatorial position of each fruit, and the regional mean spectrum was used as the spectral feature of one fruit sample. The sample-set partitioning based on joint X–Y distances (SPXY) algorithm was used to divide the calibration and prediction sets at a 3:1 ratio after outlier removal via a residual-threshold method. Four preprocessing methods, namely LOESS smoothing, standardization, vector normalization, and Savitzky–Golay (SG) smoothing, were compared. Competitive adaptive reweighted sampling (CARS) was performed with 50 Monte-Carlo sampling runs, a maximum of 30 principal components, and 10-fold cross-validation, yielding 99 characteristic wavelengths. Partial least squares regression (PLSR), support vector regression (SVR), random forest (RF), and artificial neural network (ANN) models were then established using identical input variables and sample partitions. Model performance was evaluated using the coefficient of determination for calibration ($R_c^2$), coefficient of determination for prediction ($R_P^2$), root-mean-square error of calibration (RMSEC), root-mean-square error of prediction (RMSEP), relative prediction deviation (RPD), and ratio of performance to interquartile distance (RPIQ). Under the static laboratory acquisition conditions in this work, the SSC model achieved the best prediction performance at 110 cm with SG smoothing ($R_P^2$) = 0.8949, RPD = 3.0633, RPIQ = 5.8661), whereas the firmness model obtained optimal prediction performance at 140 cm with standardization ($R_P^2$) = 0.7460, RPD = 1.9425, RPIQ = 3.2867). Changes in detection distance altered illumination uniformity, effective reflected signal, photon-scattering paths, and background-noise proportion. These effects may partially explain why the chemical-absorption-dominated SSC index and the tissue-scattering-dominated firmness index responded differently to detection distance. The results provide a reference for setting spectral detection parameters for Korla fragrant pears; however, samples were obtained from only a single producing region, harvest season, and maturity stage, and no independent external validation dataset was used. Therefore, the generalization ability of the developed models needs to be further verified using cross-season and cross-orchard sample sets. Full article

Existing reviews on AI in tea production are either agriculture-generic or limited to isolated tasks. This review thoroughly compares vision technologies (RGB, hyperspectral, near-infrared, thermal, Light Detection and Ranging (LiDAR), Unmanned Aerial Vehicle (UAV)) and establishes a task-oriented algorithm selection framework for the tea industry. For small-sample or near-linear problems, traditional machine learning (ML) (support vector machine (SVM); partial least squares regression (PLSR)) remains effective. For unstructured field tasks, deep learning achieves superior performance: pest detection accuracy exceeds 97%, tea bud detection reaches 96.8% with RGB images, and hyperspectral imaging predicts nitrogen content with R² > 0.90 and tea polyphenols with R² up to 0.925. Algorithm choice further differentiates by task granularity: lightweight convolutional neural networks (CNNs) balance speed and accuracy for edge deployment at 16 fps; You Only Look Once (YOLO) series detectors enable real-time localization on mobile platforms at 93.1% accuracy, 24 ms per target. No single algorithm dominates all tea tasks; selection is a trade-off among accuracy, speed, data availability, and computational constraints. These findings outline a structured analysis of the challenges and pathways for transitioning computer vision (CV) from laboratory research toward field-deployable tools. Full article

Accurate estimation of canopy water content (OWC) is essential for precision irrigation, crop growth monitoring, and yield prediction. Unmanned aerial vehicle (UAV)-based multispectral remote sensing provides a rapid and non-destructive approach for monitoring crop water status; however, the selection of effective spectral features and appropriate machine learning algorithms for robust OWC estimation remains insufficiently investigated, particularly across multiple growing seasons. This study evaluated the potential of UAV multispectral imagery for estimating oilseed rape canopy water content using two feature selection strategies and four representative machine learning algorithms. Field experiments were conducted during two consecutive growing seasons (2023–2024 and 2024–2025). Different sowing dates, nitrogen application rates, and planting densities were used to create a broad range of canopy water conditions. UAV multispectral images were acquired at ten representative growth stages during the reproductive period, from stem elongation to physiological maturity. Fourteen vegetation indices (VIs) were extracted from the multispectral imagery. Pearson correlation analysis and principal component analysis (PCA) were used to select informative features. These features were then used to develop multiple linear regression (MLR), partial least squares (PLS), support vector machine (SVM), and random forest (RF) models. Model performance was evaluated using each single-year dataset and the combined two-year dataset to assess robustness under different seasonal conditions. The RF model consistently achieved the highest prediction accuracy. The correlation-based RF model developed from the combined two-year dataset produced the best performance. It achieved an R² of 0.966, an RMSE of 1.734%, and an RRMSE of 2.360% for the training dataset. For the independent testing dataset, the corresponding values were 0.901, 2.794%, and 3.830%, respectively. The PCA-based models showed similar performance and effectively reduced feature redundancy. However, they did not consistently outperform the correlation-based models. These results indicate that combining UAV multispectral imagery with appropriate feature selection and machine learning algorithms can accurately estimate oilseed rape canopy water content under field conditions. Integrating data from multiple growing seasons further improves model robustness and provides a practical basis for UAV-assisted crop water monitoring and precision agricultural management. Full article

The concentrations of DIN and DIP are important indicators in an offshore ecosystem; although they do not have optical activity, their concentrations are affected by optically active substances, such as sediment, chlorophyll, and dissolved organic matter, an association that is especially close in coastal waters. This study aimed to identify this relationship to provide a theoretical basis for using remote sensing to monitor DIN/DIP concentrations. This study first used correlation analysis to analyze the relationship between water quality indicators and the field-measured spectrum in the Sheyang estuary. The results show a strong positive correlation between the DIN and DIP concentrations and spectrum in near-infrared range, similar to that between the suspended sediment concentrations and spectrum; this indicates a close relationship between DIN/DIP concentrations and sediment concentration in this sea area. Traditional regression models for DIN and DIP concentrations were constructed using the sensitive bank factors of a Hyperion image. By comparing the physical meaning of the factors and the precision and stability of the models, the quadratic model established by the ratio factor of 45th and 10th bands was selected as the DIN concentration inversion model, the quadratic model established by the ratio factor of the 45th and 9th bands was selected as the DIP inversion model, and the inversion results of the image conformed to the actual distribution pattern of DIN and DIP concentrations. In order to fully utilize the spectral information of the Hyperion data, the model coupled using partial least squares (PLS) and support vector machine (SVM) was used to construct regression models of DIN and DIP concentrations. By comparing the standardized coefficients of PLS regression, the 8~16th bands and 37~57th bands of the Hyperion image were selected; all these bands were extracted as two orthogonal components to construct the SVM regression model. Finally, the parameter combinations of radial basis model with C = 10, γ = 0.05, and ε = 0.1 and C = 1, γ = 0.1, and ε = 0.001 were determined as the inversion models for DIN and DIP concentrations, respectively. The prediction accuracy of the models was significantly improved compared to the traditional regression models, and the inversion results were superior to those of the traditional regression models, demonstrating the potential of this algorithm in hyperspectral image modeling. Full article

Artificial neural networks (ANNs) make predictions based on patterns learned during training; however, their reliability may deteriorate when the data distribution or the trained model changes. This paper proposes an Embedded One-Class Classification (EOCC) framework for monitoring task-informed neural network representations. The predictive network is first trained using its original classification or regression objective and then fixed. Embedded Support Vector Data Description (ESVDD) or Embedded Least Squares Support Vector Data Description (ELS-SVDD) is subsequently fitted to embeddings extracted from a selected hidden layer. The framework is evaluated through classification and regression simulations involving mean, covariance, and mixed distributional shifts. Additional experiments examine embedding layer and activation choices, direct neural network weight perturbations, and model changes induced by altered training conditions. Comparisons with depth-based, density-based, covariance-based, isolation-based, and end-to-end deep one-class methods show that EOCC is competitive and frequently achieves low Type II error while maintaining the nominal in-control acceptance probability, although no method is uniformly superior across all settings. Illustrative applications involving the Internet Firewall, ELEC2, and SINE1 datasets demonstrate how the framework can identify changes reflected in neural network representations. The present framework performs change detection only; adaptation and automatic model updating remain directions for future research. Full article

The nondestructive detection of sodium chloride in mural plaster layers is important for assessing salt-related deterioration in cultural heritage materials. However, the weak and indirect spectral response of sodium chloride makes accurate hyperspectral detection challenging. This study developed a hyperspectral regression framework centered on Sparrow Search Algorithm (SSA) optimization, in which continuous wavelet transform (CWT) was used to construct multiscale spectral representations and Pearson correlation analysis combined with the Successive Projections Algorithm (PCC-SPA) was used for compact variable selection. Partial least squares regression (PLSR), support vector regression (SVR), extreme gradient boosting (XGBoost), SSA-optimized SVR, and SSA-optimized XGBoost were evaluated under nested stratified specimen-grouped five-fold cross-validation. Feature selection and hyperparameter optimization were independently performed within each outer training fold, whereas the held-out specimens were reserved for performance evaluation. SVR-SSA maintained high predictive capability across both conventional and multiscale spectral representations. SG + SNV yielded an R² of 0.8167 ± 0.0690 and an RMSE of 0.3701 ± 0.0636 percentage points. Scale 6 CWT achieved closely comparable R² and RMSE values of 0.8100 ± 0.0812 and 0.3731 ± 0.0767 percentage points, respectively, together with a lower MAE of 0.2788 ± 0.0620 percentage points. Among the ten CWT scales, Scale 6 achieved the highest mean prediction accuracy, whereas Scale 2 provided the best comprehensive balance between predictive accuracy and fold-to-fold stability. These results demonstrate that the effectiveness of SSA optimization depends on the input feature representation and that CWT provides scale-resolved information beyond a single conventional spectral representation. The proposed framework provides methodological support for the nondestructive quantitative assessment of NaCl-related deterioration in mural plaster materials and establishes a basis for further application in mural conservation. Full article

Terahertz (THz) waves offer unique advantages, including non-contact operation, high penetration capability, and high resolution, making them particularly well-suited for the non-destructive thickness measurement of film-structured materials. In reflective terahertz time-domain spectroscopy (THz-TDS), thickness measurement approaches are generally classified into three categories: optimization-based methods that rely on theoretical models, time-of-flight (ToF), and machine learning. Model-based optimization techniques require precise knowledge of the optical parameters and structural configuration of each layer; however, they often suffer from slow convergence and are prone to becoming trapped in local optima. In contrast, ToF-based methods determine thickness by calculating the time delay between echo pulses reflected from different interfaces, yet their applicability is limited when the film thickness is extremely small. Machine learning, especially deep learning, enables the establishment of a direct, data-driven mapping between THz waveforms (or their extracted features) and the target thickness. Such approaches offer rapid inference, strong robustness to noise, and good adaptability to thin or structurally complex films, although their accuracy remains dependent on the quality of training data and the generalization capability of the model. In this study, high-fidelity THz waveform data generated via finite-difference time-domain (FDTD) simulations are utilized to conduct a comparative investigation into the film thickness prediction performance of several representative machine learning algorithms, including Back Propagation (BP) neural networks, Support Vector Machines (SVM), Random Forests (RF), Extreme Learning Machines (ELM), K-Nearest Neighbors (KNN), and Partial Least Squares (PLS) regression. The results indicate that, in terms of prediction error, the overall ranking of algorithmic performance from best to worst is: PLS > RF > SVM > BP > ELM > KNN. These findings provide valuable guidance for the future application of machine learning-assisted THz-TDS in precise film thickness measurement. Full article

Adulteration of freshly milled rice with rice from older sources is a fraudulent and illegal practice that exploits consumers. The purpose of this study was to develop a rapid and non-destructive technique that can detect this adulteration of milled rice, using near-infrared hyperspectral imaging (NIR-HSI) in the wavelength range of 935–1720 nm. Adulterated samples were prepared by adding old and freshly milled rice at different levels, scanning the mixed samples, and comparing the results with 100% freshly milled rice samples. All samples were divided into calibration and prediction sets for the development of classification and calibration models. Spectral pretreatment methods were tested to develop the optimum models. For qualitative prediction, the best results for differentiation between freshly milled rice and adulterated samples using partial least squares discriminant analysis (PLS-DA) yielded 96.15% accuracy, 92.86% sensitivity, and 100% specificity. For quantitative prediction, the best calibration model for determining the percentage of mixing with old rice using support vector machine regression (SVMR) yielded a coefficient of determination for prediction (R²_p) = 0.90, and root mean square errors of prediction (RMSEP) = 9.10%. These findings demonstrate the potential of NIR-HSI for both qualitative and quantitative analyses in detecting the adulteration of freshly milled rice with old rice. It can be used as a rapid, nondestructive technique for assessing the authenticity of milled rice. Full article

The accurate and rapid estimation of water content in granular materials is important for geotechnical engineering, construction-material evaluation, agricultural monitoring, and non-destructive material assessment. This study presents a controlled laboratory feasibility framework for estimating water content in sand–water mixtures using microwave S-parameter measurements and physically interpretable features. Sixty measurements were acquired from six nominal mixture levels containing 0%, 5%, 10%, 15%, 20%, and 25% water by mass using a WR-229 waveguide-based fixture over 3.30–4.90 GHz. Descriptors were extracted from the raw and empty-reference-normalized S₁₁ and S₂₁ responses, including extrema, slopes, area-based indicators, band-averaged values, selected-frequency responses, and phase statistics. Ridge regression, partial least squares regression, support vector regression, Gaussian process regression, random forest, and gradient boosting regression were evaluated. The selected PLS model achieved RMSE = 3.56%, MAE = 2.92%, and R² = 0.826 under leave-one-mixture-level-out group-wise cross-validation, which was used as the primary indicator of generalization to an unseen mixture level. The substantially lower error obtained under repeated-measurement leave-one-out cross-validation primarily reflects within-level repeatability under controlled conditions and should not be interpreted as evidence of universal calibration performance. The results demonstrate that magnitude- and phase-derived microwave descriptors, particularly transmission-based features, provide an interpretable and repeatable framework for water-content estimation under the specific sand type, sample geometry, water source, and laboratory conditions investigated. Further validation using independent preparation batches, different granular materials, measured water conductivity, temperature variation, and field-like conditions is required before practical deployment. Full article

Traditional approaches for monitoring aquatic organic carbon, such as satellite remote sensing and automated underwater sensors, are often constrained by limited temporal resolution, data gaps under cloudy conditions, maintenance requirements, and cost-effectiveness. To overcome these limitations, we developed and field-demonstrated a ground-based hyperspectral remote sensing system (GHRSS) for continuous, high-frequency monitoring of dissolved organic carbon (DOC) and particulate organic carbon (POC). The system is based on the above-water method and integrates three miniature hyperspectral spectrometers to measure water-surface radiance, sky radiance, and downwelling irradiance for deriving hyperspectral remote sensing reflectance ($R_{rs}$). The spectrometers cover 400–900 nm with a spectral resolution of 1 nm and support a minimum sampling interval of 10 s. The GHRSS also integrates solar power supply, 4G communication, and a microcomputer, enabling autonomous long-term deployment and wireless data transmission. Based on the GHRSS, retrieval models for DOC and POC were developed and validated using 90 paired in situ measurements collected from the Cao’e River. Empirical and machine learning methods were applied to retrieve DOC and POC from the measured $R_{rs}$ data. The empirical models showed limited retrieval performance, whereas partial least squares regression (PLSR) and support vector regression (SVR) substantially improved model accuracy. Among all models, SVR achieved the best performance on the independent test set, with $R^2=0.979$, RMSE = 0.031 mg/L, and MAE = 0.024 mg/L for DOC and $R^2=0.960$, RMSE = 0.152 mg/L, and MAE = 0.066 mg/L for POC. Using the optimal SVR models, minute-scale time series of DOC and POC were reconstructed from the GHRSS observations. The results revealed pronounced sub-daily variability in both parameters, with DOC varying relatively smoothly, whereas POC exhibited stronger short-term fluctuations and more rapid responses to hydrodynamic changes. These findings demonstrate that the GHRSS, combined with machine learning models, provides an effective and practical approach for continuous, high-frequency monitoring of riverine organic carbon dynamics. Full article

Despite an increasing interest in the pharmacology of cannabinoids from Cannabis sativa, little is known to date about their ability to cross the placenta and to be secreted into breast milk, and in this study, we sought to fill this gap. In total, 126 phytocannabinoids previously detected in Cannabis sativa were investigated for their transplacental transfer and secretion into breast milk. Placental transport was predicted using novel multiple linear regression (MLR), artificial neural network (ANN), boosted trees (BT), and support vector regression (SVR) models, based on a reference set of 84 compounds for which the placental clearance index (CI) relative to antipyrine is known. Secretion into breast milk was predicted using newly developed classification models based on soft independent modeling of class analogies (SIMCA) and One-Class Partial Least Squares (OC-PLS) algorithms. Analysis of the Q vs. Hotelling’s T² plot for the cannabinoids indicated that they are similar in their physicochemical properties to compounds empirically demonstrated to enter breast milk (“in-class”); only 7 of 126 compounds were borderline (with elevated Q but not T²); no compounds were classified as “out-of-class”. The mean predicted CI values for phytocannabinoids investigated in this study ranged from 0.4 to 0.85. It was concluded that all the cannabinoids in the studied group might cross the placenta (although their passage might be expected to be more difficult than that of antipyrine) and enter breast milk. These results should support informed risk assessment and prioritization of cannabinoids for future experimental testing. Full article

Accurate forecasting of country-level temperature anomalies is increasingly important for climate monitoring, policy planning, and environmental risk assessment. However, the trade-off between predictive performance, model complexity, and computational cost remains insufficiently explored, particularly across multiple countries using compact and interpretable feature representations. This study presents a comprehensive comparative evaluation of eight forecasting approaches for annual temperature anomaly prediction using country-level observations from the FAOSTAT Temperature Change dataset. The evaluated methods comprise a Persistence baseline, Ordinary Least Squares (OLS), Ridge regression, Support Vector Regression (SVR), Random Forest, a multilayer perceptron (MLP), and the classical time-series models ARIMA and ETS. Annual temperature anomalies were modeled using lagged observations, a temporal trend, and a trailing moving average under a temporally ordered 80/20 train–test split. Model performance was assessed using RMSE, MAE, $R^2$, per-country win-rate, computational runtime, and pairwise statistical comparisons based on the Wilcoxon signed-rank test with Holm correction. Hyperparameters were optimized through expanding-window temporal cross-validation, and an ablation study was conducted to quantify feature contributions. Results indicate that the ETS model achieved the best overall predictive performance, obtaining the lowest median RMSE (0.3388 °C), the lowest MAE (0.2792 °C), and the highest per-country win-rate (40.07%). ARIMA provided competitive forecasting accuracy but incurred substantially higher computational cost, whereas OLS and Ridge offered an attractive compromise between predictive performance, robustness, interpretability, and computational efficiency. In contrast, the more flexible machine learning models (SVR, Random Forest, and MLP) did not consistently outperform the simpler approaches despite their higher complexity. Overall, the results demonstrate that classical statistical forecasting methods remain highly competitive for annual country-level temperature anomaly prediction and that increasing model complexity does not necessarily translate into improved predictive performance. Full article