Azamat Serek

Built-in attention is often presented as a free explanation of what a forecasting transformer relies on, but the claim is rarely tested. We test it on MeteoFormer, a patch-based transformer with single-head variable and temporal attention that forecasts four weather variables 24 h ahead from 96 h of hourly input on the Jena and Beijing PRSA datasets. Against ten baselines trained under one protocol with five seeds, MeteoFormer has the lowest mean absolute error on Jena (3.025 against 3.107 for Crossformer, Diebold–Mariano p < 0.05 against every baseline after Holm correction) and is level with Crossformer on Beijing (4.151 against 4.138, p = 0.72), while Crossformer has the lower squared error, a difference traced to humidity. The model depends heavily on the latest observations: losing the last 16 h raises its Jena error by 73% and drops it from first to sixth place. The explanation claim fails. The fidelity of the variable attention does not differ from zero (0.012 on Jena, 0.046 on Beijing), far below SHAP on the same model; the temporal attention is almost uniform although the forecast depends mainly on the last 16 h; and the configuration selected for accuracy has the least faithful attention (0.012 against 0.181 with instance normalization, p = 0.001), but the more faithful configuration is less accurate (Jena MAE 3.195 against 3.025) and its attention remains below SHAP.

Built-in attention is often presented as a free explanation of what a forecasting transformer relies on, but the claim is rarely tested against the model's behaviour. We test it on MeteoFormer, a patch-based transformer with single-head variable and temporal attention that forecasts temperature, relative humidity, pressure and wind speed 24 h ahead from 96 h of hourly input on the Jena climate and Beijing PRSA datasets. Against ten baselines trained under one protocol with five seeds, MeteoFormer has the lowest mean absolute error on Jena (3.025 against 3.107 for the runner-up Crossformer, Diebold-Mariano p < 0.05 against every baseline after Holm correction) and the most accurate temperature forecast on both datasets, while on Beijing it is level with Crossformer in absolute error (4.151 against 4.138, p = 0.72) and Crossformer has the lower squared error, a difference that we trace to the humidity channel. The explanation claim fails. On equal budgets across all eleven models, the fidelity of the built-in variable attention does not differ from zero (0.012 on Jena, 0.046 on Beijing) and is far below post-hoc SHAP on the same model, the temporal attention is almost uniform although occlusion shows that the forecast depends mainly on the last 16 h, and the configuration selected on validation for accuracy has the least faithful attention (0.012 against 0.181 with instance normalization on Jena, p = 0.001).

Most benchmarks used to evaluate retrieval-augmented generation (RAG) systems assume a static world: they say little about how well a system handles new information, conflicting evidence, multiple languages, or questions that simply cannot be answered from the available evidence. This study introduces VERITAS, an adaptive, temporal- and conflict-aware RAG pipeline built to improve multilingual answer quality without sacrificing source grounding. It combines need-based update routing, hybrid BM25/dense retrieval, cross-lingual evidence access, multilingual reranking, temporal and source-trust ranking, conflict verification, history-aware filtering, adaptive context construction, and citation verification, all built around a single frozen Qwen3-8B generator shared across every method compared. To evaluate it, we built Adaptive Knowledge v6, a 3,000-item benchmark spanning Kazakh, Russian, and English across five knowledge conditions: stable, new, updated, conflicting, and unanswerable. On a deterministic 999-item test split, VERITAS reached 60.06% accuracy and 74.02% token F1, against 52.95% and 70.62% for the strongest baseline, a statistically significant +7.11 percentage-point gain (95% paired-bootstrap CI [4.40, 9.71]; McNemar p = 2.22 × 10⁻⁷). The improvement was largest on conflicting and unanswerable items but did not carry over uniformly to updated knowledge. VERITAS also showed a coverage–grounding trade-off, with accepted-answer coverage of 70.67%. Together, these results point to real gains under heterogeneous knowledge conditions, while also surfacing wrong abstention and updated-knowledge handling as the clearest targets for future work.

Accurate and transparent classification of agricultural land cover from satellite imagery is essential for operational crop monitoring. This study presents MS-AttNet, a compact convolutional network with squeeze-and-excitation channel attention for Sentinel-2 imagery using all thirteen spectral bands. On the EuroSAT benchmark, evaluated with a stratified 70/15/15 split across five random seeds, the model achieves 98.72 ± 0.21% accuracy, a macro F1-score of 0.9868 ± 0.0021, and a mean one-vs-rest AUC of 0.9999 with 1.32 million parameters. A ResNet-50 baseline re-implemented on the identical split achieves 97.90% accuracy with 23.56 million parameters, making the proposed model approximately eighteen times smaller and nearly three times faster at inference. A controlled ablation using five seeds per configuration shows that squeeze-and-excitation modules, convolutional stage depth, and vegetation indices as additional input channels each change accuracy by less than run-to-run variability. This indicates that the benchmark is saturated with respect to architectural choices in this model family; therefore, performance is not attributed specifically to the attention mechanism. Interpretability is assessed using Grad-CAM and SHAP band attributions with quantitative validation. Against a spatially smooth random control, Grad-CAM reduces deletion AUC from 0.857 to 0.738 and increases insertion AUC from 0.864 to 0.916 across 300 test patches. SHAP identifies blue, near-infrared, green, red, and short-wave-infrared bands as dominant contributors, with stable ordering across independently trained models. NDVI and NDWI were used only for spectral-signature analysis and interpretation, not as classifier inputs. The combination of compact size, competitive accuracy, and quantitatively verified explanations supports large-area operational monitoring applications

The rapid migration of financial fraud to encrypted messenger platforms such as WhatsApp and Telegram has created an urgent need for lightweight, endpoint-based detection tools capable of operating on small, multilingual datasets characteristic of real-world "cold start" conditions. Unlike traditional email spam, messenger-based fraud is short-lived, conversational, and heavily obfuscated through mixed scripts and artificial urgency, rendering static rule-based filters ineffective. Existing machine learning solutions predominantly address English-language SMS spam and rely on large labeled corpora, leaving a significant gap for post-Soviet and Central Asian deployment contexts where Russian, Kazakh, and English frequently co-occur within a single message. This study aims to identify the optimal classical machine learning architecture for binary fraud classification of such multilingual text messages under severe data scarcity constraints. A balanced pilot dataset of 189 messages (95 legitimate / 94 fraudulent) was manually collected from Telegram, WhatsApp, Reddit, Pikabu, and verified news sources. Text was vectorized using TF-IDF (unigrams and bigrams, max 100 features, min_df=2) without morphological preprocessing to preserve cross-language token integrity. Five classifiers were benchmarked: Logistic Regression, Linear SVM, Multinomial Naive Bayes, Random Forest, and XGBoost. Regularization strategies including L2 penalty, additive smoothing, and structural tree constraints, combined with class-weight balancing, were applied to mitigate overfitting on the small sample. Model evaluation used 5-fold stratified cross-validation and an independent test set, with Recall (Fraud) as the primary safety metric alongside Accuracy, F1-score, and ROC-AUC. Logistic Regression and Linear SVM achieved the highest overall performance (Accuracy = 0.97, Recall = 1.00, F1 = 0.97, ROC-AUC = 0.957–0.962), successfully identifying all fraudulent messages with only one false alarm each. XGBoost also reached perfect recall (1.00) at a cost of lower precision (Precision = 0.88), while Random Forest and Naive Bayes were less stable, achieving accuracies of 0.86 and 0.79 respectively. The findings demonstrate that for small, sparse, and multilingual fraud datasets, aggressively regularized linear models with minimal feature spaces outperform complex ensemble architectures, offering a viable, computationally lightweight foundation for real-time endpoint fraud filtering on encrypted platforms.

Climate change is increasing the frequency of hydrological extremes such as floods, requiring more effective approaches to build community preparedness alongside traditional flood risk management. This study develops and evaluates a serious-game-based flood risk education intervention in Kazakhstan and examines whether post-intervention protective behavioral intention can be predicted from gameplay, hydrological-literacy, and governance-attitude data. A cross-sectional, post-intervention survey of 60 participants from flood-, GLOF-, and drought-exposed regions of Kazakhstan was analyzed using descriptive/inferential statistics and three machine-learning classifiers (logistic regression, random forest, multilayer perceptron) to predict protective behavioral intention and to identify its strongest correlates. The participants showed good knowledge about immediate responses in flood events: 91.7% were able to identify the appropriate evacuation procedures, and the mean hydrological-literacy score was 5.13 out of 7. There was relatively less knowledge about longer-term mitigation measures, especially the cost effectiveness of nature-based flood protection (56.7%). Most participants agreed with the introduction of measures for flood management using nature (58.6%) and mandatory disaster insurance (87.7%), and 81.0% indicated that they planned to take preparedness measures after playing. The predictive models with the highest discrimination were logistic regression and random forest, with an area under the receiver operating characteristic curve (AUC) of 0.81 each, while the multilayer perceptron (MLP) had an AUC of 0.75. Importantly, both machine-learning and conventional statistical analyses revealed that perceived scenario fidelity was the strongest and most consistent predictor of protective behavioral intention. These results suggest that serious-game-based flood risk education provides a useful context for examining hydrological knowledge and generating behavioral data for predictive modeling of preparedness. The framework combines experiential learning, flood-governance assessment, and machine-learning-based behavioral prediction, while the findings indicate an association between perceived scenario realism and preparedness-related behavioral intention. These results provide a basis for further investigation of data-driven disaster risk education in Kazakhstan and other settings facing similar hydrological hazard risks.

Large ensembles of atmospheric dispersion simulations are necessary for uncertainty quantification, source attribution, and emergency response, however, physics-based models like HYSPLIT might be too computationally expensive for operational use. In addition, they have not fully benefited from recent advances in artificial-intelligence-based surrogate modelling, which offer the potential to reduce this computational burden. This study introduces the Dual-Head Mass-Consistent Plume Surrogate (DM-CPS), a machine-learning surrogate to promote the prediction of PM_2.5 transport at a much higher speed without compromising the geometry and integrated pollutant mass of the plume. The proposed framework utilizes two types of complementary histogram gradient-boosting models: one trained in the logarithmic concentration domain to reconstruct the plume morphology, and one trained in the original concentration domain to estimate the total pollutant burden. The outputs of the two branches are then combined through a field-level mass-consistency scaling process that preserves the spatial structure of the plume while enforcing a physically consistent integrated pollutant mass. The model is trained on a GDAS-driven ensemble of 96 HYSPLIT simulations over Almaty, Kazakhstan, which includes 1.88 million grid-cell records and physics-informed wind-aligned features. DM-CPS achieves the highest operational dispersion skill (FAC2 = 0.57) with a significant reduction in geometric bias (MG: 2.75 → 1.68) and an improvement in the median integrated-mass ratio (0.13 → 0.23) while maintaining plume structure (SSIM ≈ 0.83) compared to random forest, multilayer perceptron, and conventional gradient-boosted baselines. The surrogate predicts an entire concentration field in about 35 ms, a factor up to five times faster than running HYSPLIT. The cross-regime evaluation also shows a significant gap of generalization in the presence of unseen, out-of-training meteorological data, again emphasizing the need for increased diversity of meteorological training data for operational use. The results show that physically informed surrogate modeling can significantly speed up the prediction of atmospheric dispersion, which is fundamental to run large ensembles.

Urban tree mapping is necessary in environmental sustainability and climate change mitigation, and it depends heavily on the individual tree recognition and canopy segmentation to analyze city green cover. This systematic review discusses recent developments in the 2014–2026 mapping of these trees with the use of UAVs and high-resolution satellite imagery. Our preliminary selection of 4148 records reduced to 101 eligible publications following a systematic screening and synthesis of the records, assessed the efficiency of deep learning models such as Convolutional Neural Networks and Vision Transformers in processing various source images. We also compare object detection and semantic segmentation to see which one is more competent to deal with typical urban challenges, including overlapped canopies and building shadows. According to the reviewed studies, UAV-based models generally achieve higher spatial accuracy than satellite-based approaches for individual tree detection and crown delineation, with reported average Intersection over Union (IoU) values of approximately 70–75%, whereas satellite imagery provides superior spatial coverage for large-scale urban forest monitoring. Lastly, we present a research roadmap to address the existing weaknesses such as geographic bias, which propels the research direction towards multimodal data fusion and Foundation Models to sustain consistent, large-scale urban forest monitoring.

Accurate change detection in high-resolution remote sensing imagery is essential for urban planning, land-use monitoring, and disaster response. This study introduces CTA-Net, a Cross-Temporal Attention Network for binary change detection in bi-temporal optical imagery, designed to improve robustness against pseudo-changes caused by illumination variation, seasonal effects, and sensor noise. The proposed method employs a shared Siamese encoder with multi-scale Cross-Temporal Attention modules that derive spatial and channel attention from L2 feature differences, along with a lightweight confidence estimation head for per-pixel uncertainty modelling. A hybrid loss function combining confidence-weighted binary cross-entropy and focal loss is used to address class imbalance. Experiments on the LEVIR-CD dataset demonstrate that CTA-Net achieves an overall accuracy of 98.99%, an F1-score of 87.68%, an Intersection over Union of 78.06%, a Cohen’s kappa of 0.8715, and a Matthews Correlation Coefficient of 0.8721, with stable convergence and minimal overfitting. Qualitative and calibration analyses further indicate that the model produces interpretable attention maps and reliable probabilistic outputs. To evaluate cross-domain generalization, we conduct a transfer learning case study on multispectral Sentinel-2 agricultural imagery. The model is adapted to 11-channel input and fine-tuned on automatically generated change masks derived from NDVI-delta thresholding. Under this supervision protocol, CTA-Net achieves an F1-score of 95.18% and an IoU of 90.81% on a held-out test region, with balanced precision and recall. While these results demonstrate effective adaptation across sensor modality, spatial resolution, and semantic domain, the evaluation reflects agreement with the mask generation procedure rather than independently annotated ground truth. While CTA-Net shows strong performance and reasonable interpretability, its cross-domain evaluation is limited by the use of automatically generated labels. As a result, the reported transferability should be interpreted cautiously until validated on human-annotated datasets.

The hydrological prediction in transboundary river basins is difficult because of their heterogeneous data distribution, restrictions of data sovereignty, and the irregular nature of the discharge pattern. In this paper, Federated Learning (FL) with an LSTM neural network is proposed for next-day predictions of discharge in the Chu–Talas transboundary basin. The basin area belongs to both Kazakhstan and Kyrgyzstan. In an FL scenario, two hydrological stations from the basin were selected as client nodes, representing two different discharge regimes. Station 15125—the Chu main channel is characterized by the highest discharge regime among stations located in the basin, while Station 15233—Merke tributary represents a small catchment with an irregular regime. The federated LSTM model is compared against a centralized LSTM and a local-only LSTM baseline model. The training process is based on nearly three decades of daily discharge measurements. The preprocessing step includes synchronization, lag calculation, and windowing operation. The models are trained using three metrics: root mean square error, mean absolute error, and Nash–Sutcliffe Efficiency, as well as using Monte Carlo Dropout for estimation of the probabilistic uncertainty. The results demonstrate that the federated model demonstrates comparable performance with the centralized one for the Chu main channel. It also improves prediction accuracy for the smaller Merke tributary compared with both centralized and local-only models. These findings show that FL can work effectively with non-IID and heterogeneous hydrological data. The study makes three main contributions: (i) it implements the FedAvg algorithm on transboundary, heterogeneous hydrological data, proving that decentralized optimization can effectively capture autoregressive temporal hydrology without data centralization; (ii) it systematically compares federated, centralized, and local-only models, demonstrating that the federated approach eliminates the scale bias that traditionally neglects smaller, high-variance catchments; and (iii) it utilizes Monte Carlo Dropout to translate deterministic AI outputs into risk-aware probabilistic bounds. Ultimately, the results of this study demonstrate the practical and scientific usefulness of FL in operational water management, as the method presents a privacy-saving means of increasing predictive capacity and enabling risk-based decision-making in transboundary river basins.