The study focuses on the bearing strength of subgrade soils strengthened with geosynthetic materials, which are used in road construction to improve weak soils through functions including separation, reinforcement, filtration and drainage. CBR is a key measure of subgrade bearing capacity and informs pavement design. Conventional soaked CBR testing can be slow and labor-intensive: specimens remain submerged for 96 hours, while Proctor compaction tests to determine optimum moisture content and maximum dry density can extend the work on an individual sample to five or six days. Heterogeneous site conditions may require testing many samples. The compiled dataset therefore offers a potential route to estimating CBR from properties that can be obtained more readily, although the study does not establish that the models can replace laboratory testing in routine practice.

To build the database, the researchers combined 90 soaked CBR observations from nine previously published studies. The records covered sandy and fine-grained soils classified under the Unified Soil Classification System, with samples representing several combinations of soil properties, geosynthetic materials and reinforcement configurations. Eleven predictors were used: liquid limit, plastic limit, plasticity index, maximum dry unit weight, optimum moisture content, percentage of fines, percentage of sand, geosynthetic tensile strength, number of reinforcement layers, position of the first layer and position of subsequent layers. Soaked CBR was the target variable. Across the dataset, CBR ranged from 1.32% to 88.3%, with a mean of 16.067%. Reinforcement tensile strength ranged from 0.75 to 140, and the number of layers ranged from one to three.

The model comparison used a common 80/20 partition, retaining 72 records for training and holding out 18 for final testing. Input scaling parameters were derived only from the training data and then applied unchanged to the test data, limiting information leakage. The authors evaluated plain and AOA-assisted versions of three learners under the same split: ANN, Ridge regression and XGBoost. AOA is a population-based metaheuristic inspired by the hunting behavior of Aquila eagles. It was used to tune model weights and hyperparameters, with mean squared error serving as the optimization objective for the neural-network training. The AOA configuration used a population of 150 and up to 500 generations, and the ANN had an 11–6–1 architecture: 11 inputs, six hidden neurons and one output neuron.

AOA improved the reported test performance of every base model, although the size of the gain differed. For ANN, R² increased from 0.929 to 0.945, MSE declined from 3.06 to 2.87 and MAE fell from 18.09 to 13.87. Ridge remained the weakest performer despite optimization: AOA-Ridge reached R² = 0.831, MSE = 4.90 and MAE = 42.81, compared with R² = 0.815, MSE = 5.17 and MAE = 47.09 for plain Ridge. The AOA-XGBoost model ranked first under the study’s weighted color-intensity scheme, which assigned weights of 0.5 to R², 0.3 to MSE and 0.2 to MAE. Its advantage was especially visible in MAE, which fell by 57.1% relative to plain XGBoost. For this dataset and modeling setup, AOA-assisted XGBoost was the most accurate of the six evaluated models.

Five-fold cross-validation on the training data was used to test whether the ranking depended on the particular hold-out split. AOA-XGBoost again performed best, with cross-validated R² = 0.8819 ± 0.0765, root mean squared error of 6.4218 ± 3.1846, MAE of 3.5217 ± 1.4826, mean absolute percentage error of 19.6842 ± 4.7315%, and A20 = 0.6427 ± 0.1186. The A20 metric represents the proportion of predictions falling within ±20% of the corresponding measured values. AOA-ANN ranked second, with R² = 0.8347 ± 0.1062, while AOA-Ridge had R² = 0.6487 ± 0.3628. The broadly consistent ordering between the held-out test and cross-validation results supports the reported advantage of the optimized tree-based model, but the cross-validated R² is substantially lower than the 0.980 obtained on the test set, illustrating the sensitivity of performance estimates in a small dataset.

Training behavior for the AOA-ANN was also monitored over 500 generations. Its mean squared error fell from approximately 177 initially to around 95 in the first 20 epochs, then declined in stages to approximately 58 at epoch 40, 50 near epoch 160 and 44 around epoch 190. Improvement slowed after roughly 200 epochs, reaching about 40 by epoch 250 and approximately 38 at the end. The authors interpreted this as convergence toward a near-stable solution, with further optimization offering limited gains. The study also published the final ANN weights and biases and provided a worked calculation for a non-plastic reinforced soil condition, yielding a predicted CBR of 41.35%. That example lies within the observed CBR range and was presented as an implementation demonstration, not as independent validation of the model outside the compiled data.

Two sensitivity methods, the Cosine Amplitude Method (CAM) and Garson’s algorithm, were used to estimate the relative influence of the 11 inputs. Their combined analysis assigned 29.4% of the sensitivity to the number of reinforcement layers and 21.8% to geosynthetic tensile strength. The position of the first reinforcement layer contributed 12.1%, followed by the position of subsequent layers at 9.9% and plastic limit at 7.2%; the other variables together accounted for 19.5%. Thus, the four reinforcement-related variables collectively represented more than 73% of the model’s reported sensitivity. CAM and Garson ranked some secondary variables differently, however, and a separate stability analysis gave dry unit weight the most consistent rank even as the two methods disagreed about the ordering of tensile strength, sand and fines. The percentages are model-based importance measures rather than estimates of causal effects or direct physical contributions to strength.

The most important limitation is the database’s size and composition. Only 90 records were assembled, and differences among source studies could not be eliminated in soil type, material characteristics, specimen preparation or testing conditions. Five-fold validation and the independent hold-out set reduce but do not remove the risk of overfitting, especially for a model with 11 inputs and relatively few observations. The models should therefore be used primarily within the parameter domain represented by the compiled data; predictions outside the reported ranges involve extrapolation. The neural network and XGBoost also retain elements of a black box, and the sensitivity results do not replace direct testing or mechanistic evaluation of soil–reinforcement interaction. The study’s strongest conclusion is therefore bounded but promising: within its experimental dataset and evaluation protocol, AOA-XGBoost was the most accurate and consistently top-ranked model for estimating soaked CBR, and larger independent validation will be needed before its performance can be established beyond the current scope.

Subject of Research: Optimization-assisted machine learning for predicting the California Bearing Ratio of geosynthetic-reinforced subgrade soil

Article Title: New hybrid models integrating the aquila optimizer with machine learning algorithms to predict the strength of geosynthetic-reinforced subgrade soil

Article References: Tiang, J.-J., Ateş, B., Eirgash, M. A., Zhao, Y., Anand, P., Sharma, A., & Lim, W. H. (2026). New hybrid models integrating the aquila optimizer with machine learning algorithms to predict the strength of geosynthetic-reinforced subgrade soil. Frontiers in Built Environment, 12, Article 1947550. https://doi.org/10.3389/fbuil.2026.1947550

Image Credits: AI Generated

DOI: 10.3389/fbuil.2026.1947550

Keywords: geotechnical engineering, machine learning, XGBoost, metaheuristic optimization, California Bearing Ratio, geosynthetics, subgrade soil, sensitivity analysis, pavement design, hybrid models, hybrid, models

The challenge the researchers set out to address is structural. Exchange-rate returns may exhibit linear dependence, nonlinear relationships, regime-switching behavior, and volatility clustering, often simultaneously. Classical linear models such as ARIMA capture short-term linear dynamics but collapse when data turn complex. Nonlinear and machine learning methods handle intricate patterns but their superiority is not universal, and gains often prove sample-specific. The authors therefore built an Adaptive Time-Varying Weighted Hybrid Forecasting framework, abbreviated ATW-HyF, that combines three complementary paradigms: ARIMA for linear temporal dependence, a two-regime SETAR model for threshold-driven regime switching, and a feedforward artificial neural network for nonlinear approximation.

The genuinely novel element lies in how the framework weights its components. Conventional hybrid models rely on fixed or equal weights, implicitly assuming that model performance remains stable over time, an assumption rarely satisfied in foreign exchange markets. ATW-HyF instead updates inverse-RMSE weights recursively over a rolling window of twenty forecast errors. Models that recently produced lower errors receive larger performance scores and greater influence in the convex combination, while unreliable components are demoted automatically. The mechanism is grounded in the classical theory of optimal forecast combination, which holds that combining multiple forecasts under the minimum-variance principle reduces forecast uncertainty relative to any individual model.

The framework comes in two versions. ATW-HyF A produces adaptive point forecasts by dynamically combining the one-step-ahead predictions of ARIMA, SETAR, and ANN. ATW-HyF B extends this architecture with a GARCH(1,1) volatility layer that models the conditional variance of the hybrid forecast residuals, capturing volatility clustering and time-varying forecast uncertainty. The result is a unified mean-variance forecasting system that jointly predicts both the direction of exchange-rate returns and the evolving uncertainty surrounding them, information that matters for risk management applications such as Value-at-Risk and Expected Shortfall.

Evaluation was rigorous. The researchers employed an expanding rolling-origin, or walk-forward, procedure in which models are re-estimated at every forecast origin using only past observations, eliminating look-ahead bias. Prices were transformed into continuously compounded log returns to stabilize variance and remove nonstationarity. Performance was measured with RMSE, MAE, MAPE, sMAPE, MASE, and Theil’s inequality coefficient, while pairwise comparisons relied on the Diebold-Mariano test corrected for small samples using the Harvey-Leybourne-Newbold adjustment. The computational scale was substantial: roughly 2,510 recursive forecasting iterations per currency pair, more than 7,500 cycles and over 30,000 model estimations across three series, all implemented in Python with open-source libraries and a fixed random seed for reproducibility.

The evidence came from both simulation and real markets. In simulations built from a two-regime threshold autoregressive process with 600 observations, ATW-HyF B achieved the lowest forecast errors of all competing models, with an RMSE of 1.0140 and MAE of 0.8075, edging out the standalone ANN and the ATW-HyF A variant. This demonstrated that adaptive weighting combined with volatility modeling is particularly effective when the underlying process exhibits nonlinear and time-varying dynamics, even when volatility clustering is relatively weak.

The empirical results, however, told a more nuanced story. Using daily data from April 2016 to April 2026 with 2,610 observations per series, the study examined three currency pairs: the emerging-market USD/NGN pair and the major pairs EUR/USD and GBP/USD. Descriptive statistics revealed striking differences. USD/NGN returns displayed extreme kurtosis of 115.5 and strong positive skewness, reflecting heavy tails and policy-driven disruptions, while EUR/USD and GBP/USD were comparatively stable yet still non-normal. Jarque-Bera tests rejected normality for all series, and ADF tests confirmed stationarity, while autocorrelation functions dropped quickly to zero, indicating weak serial dependence that makes return prediction inherently difficult.

On these real series, no single model dominated. The ARIMA-ANN hybrid achieved the lowest forecast errors for USD/NGN, with an RMSE of 0.013085, and for EUR/USD, with an RMSE of 0.004809, suggesting that combining linear dynamics with nonlinear learning was sufficient for those markets. ATW-HyF B performed best for GBP/USD, recording the lowest RMSE of 0.006487. Notably, ATW-HyF A and ATW-HyF B produced effectively identical point forecast accuracy on the empirical data, implying that the GARCH volatility layer did not enhance mean forecasts, though it complemented the framework by quantifying forecast uncertainty.

The statistical tests added a twist that elevates the study beyond simple error rankings. Although ARIMA-ANN produced the lowest average forecast errors, the HLN-corrected Diebold-Mariano comparisons showed that ATW-HyF B achieved 18 statistically significant pairwise wins across all currency pairs, while no competing model recorded more than two. ATW-HyF B significantly outperformed the random walk benchmark for all three series, and it beat SETAR everywhere. The distinction between lowest average errors and most consistent statistical superiority highlights why forecast evaluation should combine error-based measures with formal predictive accuracy tests.

The broader lesson is that forecasting performance depends on the characteristics of the underlying exchange-rate series rather than on model complexity alone. Adaptive forecast combination provides a flexible alternative to conventional fixed-weight hybrids, allowing forecasting systems to adjust to changing market conditions, while volatility modeling improves usefulness for forecasting under uncertainty. The authors position the framework as an adaptive benchmark against which computationally intensive deep learning approaches, including LSTM and Transformer architectures, can be evaluated in future work, and recommend extending validation to stock indices, cryptocurrencies, commodity prices, and crisis periods to further establish robustness and generalizability.

Subject of Research: Machine learning and financial forecasting

Article Title: Data-driven adaptive hybrid models for exchange rate return forecasting

Article References: Adesina, O. S., & Obokoh, L. O. (2026). Data-driven adaptive hybrid models for exchange rate return forecasting. Frontiers in Big Data, 9, Article 1917723. https://doi.org/10.3389/fdata.2026.1917723

Image Credits: AI Generated

DOI: 10.3389/fdata.2026.1917723

Keywords: exchange rate forecasting, hybrid models, machine learning, GARCH, ARIMA, neural networks, time series, financial econometrics, Data-driven, adaptive, hybrid, models

The stakes are enormous. More than two billion people worldwide depend on groundwater as their primary drinking source, and it supplies roughly 43 percent of global agricultural water demand. In Iran, where abstraction already exceeds natural recharge, groundwater accounts for over 55 percent of total water consumption. Population growth, intensified agriculture, urbanisation, and climate change are pushing these aquifers toward declining water tables, land subsidence, saltwater intrusion, and contamination that is often irreversible. Yet traditional monitoring — accurate as it is — remains slow, expensive, and too sparse to capture the spatial patterns that determine where problems actually emerge.

The core challenge the researchers faced is fundamental: groundwater is invisible from space. Unlike lakes and rivers, whose surfaces reflect light in ways satellites can decode directly, aquifers lie buried under soil and rock. The team, led by Mozhgan Yarahmadi with Ali Haghizadeh and Leila Ghasemi, circumvented this with an indirect, surrogate-based strategy. Their reasoning was that groundwater quality is tightly coupled to what happens at the surface — land use, vegetation, soil moisture, temperature, and rainfall. If satellites can see the surface, and machine learning can learn the hidden relationships between surface signals and underground chemistry, then the orbiting sensors become proxies for what lies below.

To build those proxies, the researchers turned to Sentinel-2, the European Space Agency’s twin-satellite Copernicus mission offering thirteen spectral bands at 10 to 60 metre resolution and a five-day revisit cycle. They extracted eight bands — from blue through near-infrared — for dates matching field sampling within a ten-day window, and computed five spectral indices designed to capture chlorophyll content, turbidity, and surface moisture. These were fused with climatic variables from the ERA5-Land reanalysis dataset, including precipitation and land surface temperature. Around each of five piezometer stations, a five-kilometre buffer zone averaged the pixel values, smoothing the mismatch between point-based well measurements and satellite footprints. The result was a dataset of 56 valid samples collected seasonally from 2015 to 2024, each paired with sixteen environmental predictor variables.

With the dataset assembled, the team deployed nine competing models: five base algorithms — Random Forest, XGBoost, Support Vector Regression, Multi-Layer Perceptron, and Linear Regression — and four hybrid architectures that combined them, including a Voting Ensemble averaging four base models and a Stacking Ensemble feeding base-model outputs into a Random Forest meta-model. The standouts were striking. The Voting Ensemble achieved a coefficient of determination of 0.986 for sodium at the Sabour station and 0.974 for electrical conductivity at Gosheh, while the SVR+MLP hybrid reached 0.900 for total dissolved solids at Robat. Hybrid models claimed five of the six top-performing station-model-parameter combinations, confirming that blending complementary algorithms consistently outperforms any single learner.

The pattern behind those numbers is physically meaningful. Parameters the researchers classify as optically active — sodium, electrical conductivity, total dissolved solids, and chloride — achieved the highest accuracies, because dissolved salts alter the refractive index and light absorption of water in ways Sentinel-2’s spectrometer can detect. Semi-optical parameters such as calcium, magnesium, and total hardness landed in a moderate range, while non-optical parameters like pH, sodium percentage, and sulfate proved hardest to predict, lacking any direct spectral signature. Cross-validation told an honest story: while point-specific performance soared above 0.97, stable regional performance settled between 0.52 and 0.55 for the main parameters — respectable given the modest sample size and the inherent spatial variability of aquifer chemistry.

The study’s most novel contribution lies in how it weighted the importance of each quality parameter. Rather than relying on a single importance metric, the team combined SHAP — SHapley Additive exPlanations, a technique rooted in cooperative game theory that fairly distributes each prediction among contributing features — with classical Shapley values derived from linear regression. The final weight for each parameter was the normalised average of the two. The verdict was unambiguous: electrical conductivity dominated with 50.47 percent of the total weight, followed by total dissolved solids at 32.40 percent. Total hardness and sodium percentage ranked third and fourth at 7.87 and 7.56 percent, while the remaining six parameters each fell below half a percent. In other words, just two indicators — both optically active and strongly correlated with satellite reflectance — carry more than 82 percent of the explanatory power.

The researchers then stress-tested their rankings. Seasonal analysis revealed that while conductivity and dissolved solids held steady across spring, summer, and autumn, total hardness surged in summer — climbing from 17.12 percent in spring to 21.15 percent — likely driven by higher temperatures intensifying carbonate dissolution in the alluvial aquifer. Sodium percentage also crept upward in summer, a plausible fingerprint of agricultural fertiliser use during the growing season. Bootstrap resampling with 1,000 iterations quantified the uncertainty: conductivity and dissolved solids showed the lowest coefficients of variation, around 9.7 to 9.9 percent, confirming their weights are stable and reliable, while chloride and pH carried the highest uncertainty, consistent with their weak spectral linkage.

The authors are candid about limitations. Fifty-six samples is a small dataset, and reanalysis-based climatic data may diverge from ground stations; atmospheric correction remains tricky under cloudy skies, and models trained on Khorramabad’s geology cannot simply be transplanted elsewhere without local retraining. Yet the practical implications are compelling. Because Sentinel-2 imagery and ERA5-Land data are freely available and continuously refreshed, the trained models can be extrapolated to unmonitored areas, generating spatial maps of conductivity, dissolved solids, sodium, and chloride where no wells exist. For water managers in semi-arid regions, the framework offers an interpretable, low-cost decision-support tool: prioritise conductivity and dissolved solids as headline indicators, intensify hardness monitoring during summer, and treat satellite-based assessment as a powerful complement — not a replacement — for laboratory sampling. As aquifers worldwide face mounting pressure, watching them from orbit may soon become standard practice.

Subject of Research: Prediction and weighting of groundwater quality parameters using Sentinel-2 remote sensing, hybrid machine learning models, and SHAP-game theory analysis in Khorramabad, Iran

Article Title: Groundwater Quality Parameter Weighting Using Hybrid Machine Learning and SHAP-Game Theory Based on Sentinel-2 Data (Case Study: Khorramabad, Iran)

Article References: Yarahmadi, M., Haghizadeh, A., & Ghasemi, L. (2026). Groundwater Quality Parameter Weighting Using Hybrid Machine Learning and SHAP-Game Theory Based on Sentinel-2 Data (Case Study: Khorramabad, Iran). Environmental Advances, Article 100770. https://doi.org/10.1016/j.envadv.2026.100770

Image Credits: AI Generated

DOI: 10.1016/j.envadv.2026.100770

Keywords: groundwater quality, Sentinel-2, machine learning, SHAP, game theory, remote sensing, hybrid models, Iran, water monitoring, electrical conductivity, total dissolved solids, semi-arid regions

The classical backbone of the new method is the Enhanced Deep Super-Resolution network, or EDSR, a residual architecture introduced in 2017 that removed the batch-normalization layers of earlier designs in order to squeeze more capacity out of very deep stacks of convolutional filters. EDSR and its adversarial cousin ESRGAN set benchmarks that later attention-based and transformer-based models such as SwinIR have chased ever since. In the hybrid scheme, the authors retain this proven classical machinery but insert Variational Quantum Circuits, small parameterized quantum circuits whose rotation gates are tuned by a classical optimizer, into the learning pipeline. These circuits act as learnable nonlinear feature processors: classical image features are encoded into quantum states, transformed by entangling gates and parameterized rotations, and measured back into classical values that feed the reconstruction pathway. The approach follows a broader pattern in quantum machine learning, where hybrid models pair classical networks for heavy feature extraction with quantum subroutines that explore expressive function classes in exponentially large Hilbert spaces.

Training proceeds on paired datasets of low-resolution and high-resolution images, with the model learning to reconstruct sharp outputs from degraded inputs. Crucially, the training objective constrains both the low-resolution and high-resolution reconstructions, a consistency requirement that forces the network to maintain a coherent mapping between scales rather than merely hallucinating plausible detail at the output end. The authors trained on the DIV2K dataset, a standard corpus of two-thousand high-quality images widely used in super-resolution research, and evaluated on the Set14 benchmark, one of the canonical test sets in the field. This paired, supervised regime keeps the work firmly within the mainstream of super-resolution methodology even as the quantum component pushes into less familiar territory.

One of the more elegant aspects of the design is its treatment of perceptual quality. Many modern super-resolution systems measure perceptual fidelity using a loss computed by a large network pre-trained on an unrelated task, typically a VGG-style classifier trained on ImageNet. That strategy works, but it imports an external dependency and a computational burden. Gohel and Joshi sidestep it entirely by adopting an Inhomogeneous Gaussian Markov Random Field prior as a spatial regularizer. IGMRF models, which have a history in image restoration going back to learning-based super-resolution work by Gajjar and Joshi in 2010, capture the statistical structure of local image neighborhoods and penalize reconstructions that violate the smooth-yet-textured character of natural images. By embedding this statistical prior directly in the loss, the framework encourages visually convincing reconstructions without needing any externally pre-trained network, a simplification that matters both for training efficiency and for keeping the pipeline self-contained.

The evaluation protocol covers the three metrics that dominate the super-resolution literature. Peak Signal-to-Noise Ratio quantifies pixel-level fidelity, the Structural Similarity Index captures perceptual agreement in luminance, contrast, and structure, and Learned Perceptual Image Patch Similarity, or LPIPS, measures distance in the feature space of a deep network and correlates well with human judgments of visual quality. The authors report results for four-times and eight-times magnification, and they demonstrate a particularly practical trick: a model trained only for two-times enhancement can be applied recursively, feeding its own output back as input, to reach higher magnification factors without any architectural modification. Quantitative and qualitative examples in the paper illustrate this recursive enhancement up to eight times, suggesting that a single lightweight model could serve a range of zoom requirements rather than requiring a separately trained network for each scale factor.

Because the quantum components run on classical simulators rather than actual quantum hardware, the authors are careful about what they claim. All quantum circuits were simulated on conventional hardware, and the results are presented explicitly as evidence for the viability of the hybrid representation in a simulation-based setting, with hardware-level quantum acceleration left as a direction for future investigation. That honesty is notable in a field where quantum machine learning claims sometimes outrun the evidence. The team did, however, probe robustness in two ways that anticipate real devices. They simulated depolarizing noise, the dominant error channel on today’s noisy intermediate-scale quantum processors, and they performed circuit-connectivity ablations, testing how performance changes when qubits cannot all interact directly and gates must be routed through limited coupling graphs. Both ablations showed gradual performance changes under the tested settings rather than catastrophic collapse, an encouraging sign that the representation degrades gracefully as hardware imperfections accumulate.

The work sits within a rapidly growing literature at the intersection of quantum computing and image processing. Quantum machine learning in feature Hilbert spaces, demonstrated experimentally by Havlíček and colleagues at IBM in a landmark Nature paper, provides the theoretical footing for encoding classical data into quantum states where linear separability can be enhanced. Variational quantum circuits have been applied to deep reinforcement learning, transfer learning, and quantum generative adversarial frameworks proposed by Lloyd and Weedbrook. In imaging specifically, a parallel line of research by Dutta and collaborators has explored quantum many-body concepts for denoising, quantum-inspired convolutional networks for medical image restoration, and most recently a quantum image enhancement transformer for single-image super-resolution. The new study distinguishes itself by embedding quantum circuits inside a fully supervised, residual super-resolution pipeline with a principled statistical prior, rather than treating quantum enhancement as a bolt-on denoising stage.

The practical stakes are considerable. Super-resolution underpins satellite imagery analysis, medical imaging, video streaming, forensic enhancement, and the upscaling pipelines inside consumer displays. A hybrid architecture that achieves competitive fidelity while remaining trainable with modest external dependencies could eventually matter wherever image quality is limited by sensor resolution or bandwidth. The recursive magnification result is especially appealing for deployment scenarios where storage and compute are constrained, since one compact model could serve multiple zoom levels. Of course, the caveats are equally clear: with everything simulated classically, the framework currently offers no quantum speedup, and the overhead of simulating quantum circuits on classical hardware makes the hybrid model a research vehicle rather than a production tool. The authors frame their contribution as evidence of a promising representation, not a replacement for state-of-the-art classical systems.

That framing reflects a mature understanding of where quantum machine learning stands today. Real quantum processors remain small, noisy, and limited in connectivity, and theoretical obstacles such as barren plateaus, the vanishing-gradient problem identified by McClean and colleagues in quantum neural network training landscapes, continue to challenge the scalability of variational approaches. Studies like this one, which test robustness to depolarizing noise and restricted connectivity before hardware ever enters the picture, help map which designs are likely to survive the transition to real devices. The gradual performance degradation observed under the tested settings suggests the architecture has some tolerance for the imperfections that define the current era of quantum computing.

For now, the study stands as a careful, well-benchmarked demonstration that quantum circuits can be woven into a demanding computer vision task without sacrificing the discipline that makes super-resolution research credible: paired training data, standard metrics, public datasets, and honest accounting of what is simulated and what is not. Whether variational quantum circuits ultimately deliver advantages on actual quantum hardware remains an open question, but this work provides a concrete, reproducible template for how the marriage of quantum computing and image restoration might be tested as hardware matures. The authors received their manuscript on 22 March 2026, had it accepted on 24 September 2026, and published it on 9 October 2026 in volume 25 of Quantum Information Processing, adding a measured but intriguing data point to one of the most closely watched experiments in modern computing: teaching quantum machines to see.

Subject of Research: Hybrid quantum-classical deep learning for single-image super-resolution

Article Title: Quantum single-image super resolution

Article References: Gohel, P., & Joshi, M. (2026). Quantum single-image super resolution. Quantum Information Processing, 25(10), Article 337. https://doi.org/10.1007/s11128-026-05357-0

Image Credits: AI Generated

DOI: 10.1007/s11128-026-05357-0

Keywords: quantum machine learning, super-resolution, variational quantum circuits, EDSR, IGMRF prior, image reconstruction, deep learning, quantum computing, PSNR, SSIM, LPIPS, hybrid models

The method behind the study is as revealing as its findings. The authors used VOSviewer, a widely adopted bibliometric mapping tool, to construct keyword co-occurrence networks, run thematic clustering, and generate time-weighted overlay maps. These techniques allow researchers to visualize how scientific terms cluster together and how those clusters evolve over time. In the early years of the study window, the dominant keywords were unmistakably computational: deep learning, classification, and algorithmic approaches filled the maps. The literature of that era reads like a computer science conference, with psychology often relegated to a supporting role as a source of data rather than a partner in design.

Then came the inflection point. According to the overlay analysis, after 2023 the thematic center of gravity shifted toward human-centered concepts: large language models, digital empathy, therapeutic alliance, and personalization surged in prominence. This is not merely a change in fashionable vocabulary. The therapeutic alliance — the collaborative bond between therapist and client — is one of the most robust predictors of psychotherapy outcomes, established in decades of meta-analytic research. Its appearance as a central theme in AI research signals that developers and scholars are no longer asking whether machines can process clinical language, but whether they can participate in the relational work that makes therapy effective.

To interpret this shift, the authors turned to Everett Rogers’ Diffusion of Innovations model, a classic framework describing how new technologies spread through communities. In Rogers’ schema, adoption proceeds through stages: innovators and early adopters experiment first, followed by the early majority, late majority, and laggards. The bibliometric evidence suggests that AI in psychology has crossed a critical threshold. The early phase, dominated by technical experimentation and proof-of-concept studies, corresponds to the innovator and early adopter stages. The post-2023 explosion of work on empathy, personalization, and therapeutic relationships marks the field’s entry into the early majority stage, where the question is no longer whether the technology works but how it should be integrated into everyday practice.

The geographical distribution of the literature tells its own story. Publication output was concentrated overwhelmingly in China and the United States, with the United Kingdom, Germany, and Australia forming a second tier. This concentration mirrors broader patterns in global AI research, where the two superpowers dominate output and impact. Notably, prior research cited in the study suggests that China and the United States produce more impactful AI research when they collaborate rather than compete — a finding with real implications for mental health technology, where shared standards and cross-cultural validation could determine whether AI tools work for patients everywhere or only for some.

Perhaps the most consequential finding concerns the direction of research intensity itself. The analysis shows a clear movement away from fully automated systems and toward hybrid models in which AI augments rather than replaces human practitioners. The authors describe this as a positioning of artificial intelligence around the axis of enhancement rather than replacement. It is a reframing with enormous stakes. Surveys of the field have repeatedly asked whether AI could replace psychotherapists, and the new evidence suggests the literature has largely answered: not replace, but extend. AI systems that draft empathic responses, flag risk, personalize interventions, or train novice counselors appear to be valued precisely because they keep humans in the loop.

The empirical work behind this shift lends it credibility. Studies cited in the bibliometric map include findings that ChatGPT outperformed humans on certain emotional awareness evaluations, that human-AI collaboration produced more empathic conversations in text-based peer-to-peer mental health support, and that large language models can help upskill novice counselors through simulated practice and feedback. Other research has compared responses from human therapists and LLM-based chatbots to assess therapeutic communication quality, and integrative reviews have begun asking whether a genuinely digital therapeutic alliance exists at all. Together, these studies form the intellectual scaffolding of the post-2023 thematic cluster that the bibliometric analysis identifies.

The timing matters. The World Health Organization’s 2022 world mental health report documented a staggering global gap between need and available care, with millions unable to access any form of treatment. Conversational agents and AI-based interventions have long been promoted as a scalable answer to that shortage, and meta-analyses of AI-based conversational agents for mental health and well-being have shown promising, if still preliminary, results. The bibliometric evidence suggests the field is now grappling seriously with the conditions under which such tools can deliver not just information but genuine relational benefit — and with the ethical questions, including bias in AI mental health applications, that scale-up inevitably raises.

For clinicians, researchers, and technology developers, the study offers a roadmap of where the field has been and where it is heading. The interdisciplinary structure it documents — spanning psychology, computer science, psychiatry, and engineering — reflects a maturing consensus that no single discipline can build trustworthy mental health AI alone. The transition from algorithmic dominance to human-centered themes suggests that the next wave of innovation will be judged not by benchmark accuracy alone but by whether systems can sustain something resembling a working alliance, adapt to individual users, and enhance the capacities of human caregivers rather than simulating them wholesale.

The decade from 2015 to 2025 began with machines learning to classify and ended with researchers asking machines to care. Whether digital empathy can ever be more than a well-engineered imitation remains an open and contested question. But the bibliometric record is unambiguous about the direction of travel: the AI revolution in psychology has moved from the laboratory to the consulting room, from replacement to enhancement, and from algorithms to empathy — and the early majority is now paying attention.

Subject of Research: Bibliometric mapping of artificial intelligence research trends in psychology and mental health from 2015 to 2025

Article Title: From algorithms to empathy: a bibliometric analysis of the AI revolution in psychology and related fields (2015–2025)

Article References: Kusci, İ., & Bakır, V. (2026). From algorithms to empathy: a bibliometric analysis of the AI revolution in psychology and related fields (2015–2025). Current Psychology, 45(18), Article 1510. https://doi.org/10.1007/s12144-026-10070-3

Image Credits: AI Generated

DOI: 10.1007/s12144-026-10070-3

Keywords: artificial intelligence, psychology, bibliometric analysis, large language models, digital empathy, therapeutic alliance, mental health, ChatGPT, diffusion of innovations, psychotherapy, hybrid models, digital mental health

Deepfakes have grown dramatically more convincing and easier to produce as generative models have advanced. What once required Hollywood-grade visual effects studios can now be done on a laptop with open-source software, and manipulated videos of politicians, celebrities, and ordinary people have become a genuine threat to public trust, digital integrity, and information security. Detection systems have kept pace on paper, often reporting near-perfect accuracy on standard benchmark datasets, but that laboratory performance frequently evaporates in the real world. A detector trained on one set of manipulation techniques may be nearly blind to forgeries produced by a different generator, a phenomenon researchers call domain shift.

The new framework, developed by Abdullah Mohammad Sakib, Md. Mohashin Hossain, Mahedy Hasan Foysal, Ishtiak Al Mamoon, and colleagues at the International University of Business Agriculture and Technology and United International University in Dhaka, addresses the generalization problem with a hybrid neural architecture. The system combines two complementary types of deep learning models inside a single detector. The first branch is a convolutional neural network, or CNN, the workhorse of modern image analysis. Convolutional filters excel at picking up local, high-frequency patterns, and in the deepfake context that means subtle forensic traces: blending seams where a synthesized face meets the original head, compression artifacts left behind by the generation pipeline, and other pixel-level inconsistencies that betray manipulation.

The second branch is a Vision Transformer, or ViT, a relatively newer architecture that treats an image as a sequence of small patches and uses attention mechanisms to model relationships between distant parts of the picture. Where the CNN sees the fine texture of a boundary, the transformer captures the global semantic structure of a face: whether the eyes, nose, mouth, and surrounding regions are mutually consistent in lighting, geometry, and expression. Deepfake artifacts often manifest at both scales simultaneously, so fusing the local evidence from the CNN with the global reasoning of the ViT gives the detector a multi-scale representation that neither model could achieve alone. The authors report that this fused representation is what allows the system to remain robust when the characteristics of the data shift between sources.

The second half of the innovation lies in how the model is trained. In conventional machine learning, all data is gathered into one place and the model learns from it centrally. That approach is problematic for deepfake detection because the training material consists largely of real faces, and collecting facial imagery from hospitals, social media platforms, banks, or government agencies into a single database raises serious privacy and legal concerns. Federated learning offers an alternative: instead of sending data to the model, the model is sent to the data. Multiple clients, such as institutions or devices, each train the model locally on their own private data, and only the resulting parameter updates, not the raw images, are shared and aggregated into a global model.

Federated learning, however, comes with its own well-known difficulty. When the data held by different clients is statistically dissimilar, a situation known as non-independent and identically distributed, or non-IID, data, the local updates can pull the shared model in conflicting directions, causing unstable convergence and degraded performance. A hospital’s dataset might contain mostly one type of manipulation, while a social media company’s dataset contains entirely different forgery methods and image qualities. Standard federated optimization can struggle to reconcile these divergent signals. The researchers designed FedHybrid-ViT around this challenge, using a collaborative federated optimization protocol intended to keep training stable even under highly heterogeneous client distributions.

To test the framework rigorously, the team constructed a heterogeneous multi-source dataset from publicly available deepfake benchmarks, including FaceForensics++, a widely used collection of manipulated facial videos, and evaluated the system under both independent and non-IID distribution settings. The non-IID configuration simulates the messy reality of deployment, where each participating client sees only a narrow and skewed slice of the world’s deepfake varieties. The results showed that FedHybrid-ViT consistently outperformed standalone CNN models, standalone ViT models, and standard federated learning baselines across the test conditions.

The headline number is an area under the curve, or AUC, score of 0.925. AUC measures a detector’s ability to distinguish genuine media from fakes across all possible decision thresholds, with 1.0 representing perfect separation and 0.5 representing random guessing. A score of 0.925 indicates strong discriminative power, but the more significant finding, according to the paper, is the stability of that performance across highly heterogeneous conditions. Detectors that overfit to dataset-specific artifacts typically show sharp performance drops when evaluated on data from sources they never saw during training. The hybrid federated model maintained its accuracy under exactly those stresses, suggesting that the combination of dual-stream architecture and federated optimization genuinely improves generalization rather than merely inflating benchmark scores.

The implications extend beyond the immediate technical achievement. A privacy-preserving deepfake detector that generalizes across data sources could be deployed in federated settings where no single organization can or should aggregate facial data: networks of banks verifying customer identities, social media platforms moderating content, or forensic labs collaborating across jurisdictions. Each participant would benefit from a detector trained on the collective experience of the network without ever exposing its own users’ images. The authors argue that this makes the approach promising for scalable, distributed deepfake detection scenarios, where the diversity of manipulation techniques is precisely what defeats centralized models trained on narrow datasets.

The work also reflects a broader trend in machine learning research toward architectures that combine the complementary strengths of convolutional and transformer-based models, and toward training paradigms that respect data sovereignty. As generative models continue to improve, the arms race between forgery and detection will likely intensify, and the systems that endure may be those that can learn from many distributed sources of evidence without collapsing under the weight of their diversity. FedHybrid-ViT, with its reported AUC of 0.925 and stable behavior on non-IID data, offers one blueprint for that next generation of detectors, demonstrating that the path to robust synthetic media forensics may run not through bigger centralized datasets, but through smarter collaboration among many smaller ones.

Subject of Research: A federated hybrid CNN and Vision Transformer framework for privacy-preserving, generalizable deepfake detection on heterogeneous non-IID data

Article Title: FedHybrid-ViT: A Federated-Hybrid Framework for Generalizable Deepfake Detection on Heterogeneous Data

Article References: Sakib, A. M., Hossain, M. M., Foysal, M. H., Muzahidul Islam, A. K. M., & Mamoon, I. A. (2026). FedHybrid-ViT: A Federated-Hybrid Framework for Generalizable Deepfake Detection on Heterogeneous Data. Machine Learning, 115(10), Article 235. https://doi.org/10.1007/s10994-026-07171-2

Image Credits: AI Generated

DOI: 10.1007/s10994-026-07171-2

Keywords: deepfake detection, federated learning, vision transformer, convolutional neural network, non-IID data, domain generalization, privacy-preserving AI, hybrid models, statistical heterogeneity, machine learning, media forensics, synthetic media

Streamflow forecasting, the science of predicting how much water will pass through a river system hours, days, or even months into the future, underpins nearly every major decision in water resources. It determines when reservoir operators release water for irrigation, how hydropower plants schedule generation, and how communities brace for floods or droughts. A study of the Yellow River in China showed that a hybrid forecasting system improved hydropower generation by up to 3.39 percent, underscoring the real economic stakes involved in getting these predictions right.

The challenge lies in the sheer complexity of streamflow behavior. Unlike simple linear systems, river flows respond to a tangled web of variables including rainfall, temperature, snowmelt, and large-scale climate oscillations. Traditional statistical models such as ARIMA and its variants, while effective for smooth time series, break down when confronted with the non-linear and non-stationary character of real hydrological data. AI methods, by contrast, excel at detecting hidden patterns in noisy and multi-dimensional datasets.

The review identified four broad families of AI techniques that researchers have deployed over the past two decades. Machine learning models, including feed-forward neural networks, convolutional neural networks, recurrent neural networks, and long short-term memory networks, form the largest group. Evolutionary computation algorithms such as genetic algorithms and particle swarm optimization help fine-tune model parameters. Wavelet conjunction models use signal processing tools to decompose complex hydrological signals. Finally, hybrid models combine multiple approaches to exploit their complementary strengths.

Among the most promising recent developments is the long short-term memory network, or LSTM, which uses internal gates to selectively retain or discard information over long sequences. This architecture allows the model to capture temporal dependencies that extend far into the past, a capability that proves critical for predicting streamflow driven by slow-moving processes like snow accumulation and spring melt. Recent findings suggest that LSTMs trained on data from multiple basins can generalize to entirely new catchments, and can even forecast extreme events without having been explicitly trained on them.

Support vector machines have also proven remarkably robust, particularly in monthly streamflow forecasting where their insensitivity to outliers gives them an edge over conventional neural networks. Adaptive neuro-fuzzy inference systems, which blend neural network learning with fuzzy logic, handle uncertainty and sudden anomalies in streamflow data better than many deep learning alternatives. Extreme learning machines offer rapid computational performance, making them attractive for real-time forecasting applications.

Despite these advances, the review revealed a striking pattern: the overwhelming majority of studies have focused on short forecast windows of one to seven days ahead. Longer lead times, which would be far more valuable for reservoir planning and drought management, remain largely unexplored. Accuracy inevitably degrades as the forecast horizon extends, but the authors argue that this challenge should spur development rather than deter it. Models capable of producing reliable forecasts weeks or months ahead would transform water resources planning.

Input variable selection emerged as another critical factor shaping model performance. Most studies relied on lagged streamflow values alone, or combined streamflow with rainfall and temperature. Yet snow, a direct contributor to streamflow in mountainous basins, has been largely ignored due to data scarcity. Wind speed, which indirectly influences streamflow by shaping rainfall patterns, has been almost entirely neglected. The review suggests that incorporating these underused predictors could yield significant gains in forecast accuracy and reliability.

The analysis also found that hybrid models combining optimization algorithms with AI predictors consistently outperformed standalone approaches. Wavelet-based de-noising techniques, evolutionary algorithms for feature selection, and the fusion of deep learning architectures specialized in different tasks all contributed to more robust forecasting systems capable of handling complex, non-linear hydrological processes.

Looking forward, the review identifies several priority research directions. Time-series neural networks such as recurrent and LSTM architectures deserve far more attention than they have received. High-resolution hourly data, which could give stakeholders more lead time for flood response, remains almost entirely untapped. And as climate change intensifies the frequency and severity of extreme hydrological events, the need for AI systems capable of producing accurate, long-range streamflow forecasts has never been more urgent.

Beyond the architectural choices of individual models, the review sheds light on the practical realities that shape how these forecasting systems are built and judged. Hydrologists evaluating AI models typically rely on a small set of statistical measures, each with distinct strengths and blind spots. The mean absolute error treats all deviations equally, making it intuitive but insensitive to the magnitude of extreme failures. The root mean squared error penalizes large mistakes more heavily, which matters when a badly missed flood peak carries far greater consequences than a series of minor daily errors. The coefficient of determination and the Nash–Sutcliffe efficiency coefficient, meanwhile, gauge how well a model captures the overall shape and variability of observed flows. Because no single metric tells the whole story, the choice of evaluation criteria can materially influence which models appear superior, and the review emphasizes that researchers should match their metrics to the operational purpose of the forecast rather than defaulting to convention.

Another theme running through the reviewed literature is the role of uncertainty. Streamflow prediction is inherently probabilistic, shaped by chaotic atmospheric processes, imperfect measurements, and structural simplifications within the models themselves. The source evidence highlights that confronting these uncertainties head-on, rather than producing single deterministic numbers, is essential for achieving results that decision-makers can trust. Reservoir operators weighing a water release against an uncertain future inflow need to know not just the most likely outcome but the range of plausible alternatives. This perspective connects to the broader hydrological recognition that errors in forecasting carry real costs, and that quantifying them honestly is as important as reducing them.

The geographic and climatic diversity of the reviewed studies also matters. Snow-dominated basins behave very differently from rainfall-driven ones, and the review notes that snow-related inputs have been used only negligibly across the literature. In regions where winter snowpack functions as a natural reservoir, releasing water gradually during spring melt, models that ignore snow depth, density, and melt timing must compensate with other signals, often with limited success. Climate drivers such as the El Niño–Southern Oscillation offer another underexploited avenue, since these large-scale ocean-atmosphere patterns can teleconnect to river flows months in advance and are particularly valuable for the long-lead forecasts the review identifies as underdeveloped.

The division between short-term and long-term forecasting reflects fundamentally different physical regimes. At hourly to daily scales, streamflow is governed largely by immediate weather, making rainfall and temperature the dominant predictors, which aligns with the review’s finding that these two variables dominate model inputs. At weekly to annual scales, the relevant memory resides in snow accumulation, soil moisture stores, groundwater, and ocean-atmosphere oscillations. The concentration of research effort in the short window means the scientific community has optimized tools for flood response while leaving a comparative void around the slower, planning-oriented decisions, irrigation scheduling, reservoir storage targets, drought contingency, that depend on longer horizons.

The historical arc captured by the 1999-to-2023 window is itself instructive. Classical statistical models such as auto-regressive integrated moving average approaches established the baseline against which machine learning methods were first tested. Early artificial neural networks demonstrated that flexible, data-driven function approximators could outperform linear assumptions on non-linear hydrological series. Subsequent decades brought progressively more sophisticated architectures, alongside hybrid schemes in which genetic algorithms or particle swarm optimization tuned network weights or selected input features, and wavelet transforms stripped noise from raw signals before learning began. The consistent finding that hybrids outperform standalone models suggests that no single technique captures the full complexity of streamflow behavior, and that decomposition, optimization, and prediction are best treated as complementary stages of a single pipeline.

For practitioners, the review’s findings carry actionable weight. Selecting an appropriate model family should follow from the forecast horizon and data availability: feed-forward networks and support vector machines remain reliable workhorses for short-lead predictions, while recurrent and LSTM architectures are better suited to capturing the temporal dependencies that longer horizons demand. Investing in better input data, particularly snow observations and high-resolution hourly records, may yield larger accuracy gains than further architectural tinkering. And as computational intelligence, data mining, and time series analysis techniques continue to mature, the opportunity identified by the authors, extending reliable forecasts across larger windows and richer sets of predictors, stands to benefit irrigation districts, hydropower operators, and flood-prone communities alike, provided future research pursues the gaps this quarter-century synthesis has so clearly mapped.

Subject of Research: The use of artificial intelligence methods, including neural networks and hybrid models, for forecasting river streamflow over different lead times

Article Title: The Application of Artificial Intelligence in Streamflow Forecasting: A Review

Article References: Eswara, S., Barton, A., Choudhury, T., & Chubb, T. (2026). The Application of Artificial Intelligence in Streamflow Forecasting: A Review. Water Resources Management, 40(11), Article 522. https://doi.org/10.1007/s11269-026-04883-x

Image Credits: AI Generated

DOI: 10.1007/s11269-026-04883-x

Keywords: streamflow forecasting, artificial intelligence, neural networks, LSTM, water resources management, hydrology, machine learning, hybrid models, river flow prediction, forecast lead time, support vector machine, ANFIS

The research, led by Sharmin Akther of Jahangirnagar University together with colleagues from the Bangladesh Meteorological Department, the University of Melbourne and other institutions, focused on Mymensingh district, located at 24.75 degrees north and 90.40 degrees east. The team assembled a remarkable 38-year record of daily minimum temperatures spanning 1985 to 2022, obtained from the Bangladesh Meteorological Department. Only 30 of the 13,787 daily records, a mere 0.22 percent of the dataset, contained missing values, and these were filled using time-based interpolation that respects the temporal relationship between adjacent observations. Statistical tests confirmed the series was stationary: the Augmented Dickey-Fuller test rejected a unit root with a statistic of minus 11.426, while the KPSS test failed to reject stationarity, giving the researchers a solid foundation for time series modeling.

The core of the study was a head-to-head comparison of forecasting approaches. On the statistical side, the team fitted an autoregressive integrated moving average model, selecting ARIMA(1,1,2) as optimal using the Akaike and Bayesian information criteria, and an exponential smoothing state space model, where the simplest ETS(A,N,N) configuration with additive errors, no trend and no seasonality achieved the lowest AIC. On the machine learning side, they trained support vector regression with a radial basis function kernel, random forest regression, and a long short-term memory neural network, the deep learning architecture specifically designed to capture long-range dependencies in sequential data through its gated memory cells. Each machine learning model was fed 30 days of lagged temperatures as input features, standardized with Z-score normalization, and tuned through five-fold time series cross-validation to prevent any leakage of future information.

The team also built six hybrid models that combined each machine learning predictor with a statistical correction of its residuals, following the classic hybridization strategy in which a neural network captures nonlinear patterns while ARIMA or ETS models any remaining linear structure. The final hybrid prediction was the sum of the machine learning forecast and the statistical model’s forecast of the residuals. This family of models, including LSTM+ARIMA, LSTM+ETS, SVR+ARIMA, SVR+ETS, RF+ARIMA and RF+ETS, was evaluated with the same rigorous out-of-sample protocol applied to the standalone models.

When the models were tested on data they had never seen, covering June 2011 through September 2022, the deep learning model emerged as the clear winner. The LSTM achieved a root mean square error of 1.395 degrees Celsius, a mean absolute error of 1.053 degrees, and a mean absolute scaled error of 0.906, meaning it beat a naive persistence forecast. Support vector regression came in a close second at 1.403 degrees RMSE, and random forest followed at 1.435 degrees. The statistical models, by contrast, performed poorly, with RMSE values around 6.7 to 6.8 degrees and MASE values above 4. The researchers attribute the LSTM’s success to three factors: daily minimum temperature exhibits stronger day-to-day persistence than mean or maximum temperature, the 38-year training record provides ample data for learning seasonal cycles, and the single-station design avoids errors from spatial heterogeneity.

Perhaps the most surprising finding concerned the hybrid models. Despite the theoretical appeal of combining statistical and machine learning methods, none of the six hybrids significantly improved on its standalone counterpart. The best hybrid, LSTM+ARIMA, achieved an RMSE of 1.397 degrees, essentially identical to the standalone LSTM’s 1.395 degrees. Diebold-Mariano tests, which formally compare predictive accuracy between competing forecasters, confirmed that only LSTM+ARIMA differed significantly from its base model, and in that case the standalone LSTM was actually better. The message is that when a deep learning model already captures the complex temporal structure of a temperature series, bolting on a statistical correction adds complexity without adding skill.

Accuracy in predicting temperature, however, is not the same as accuracy in detecting cold waves, and here the study delivers its most important caution. Treating cold wave detection as a binary classification problem with the 10 degree threshold, the LSTM showed excellent discriminative power, with a ROC-AUC of 0.975, meaning it ranks cold days above ordinary days with remarkable consistency. Yet its recall was only 0.215: the model correctly identified just 14 of the 65 actual cold wave days in the test period, missing 51 of them. Precision stood at 0.560, so when the model did flag a cold day it was right 56 percent of the time, and it raised only 11 false alarms. The overall accuracy of 0.985 is misleading because cold days make up only 1.6 percent of observations, a class imbalance that pushes models toward conservative behavior. Monthly analysis revealed the pattern in detail: the model detected 22 of 43 cold days in December, a 51 percent detection rate, but only 2 of 18 in November and 1 of 4 in January, suggesting systematic underestimation of early winter cold events.

Using the trained model, the researchers generated a daily minimum temperature forecast for 2027 with 80 and 95 percent prediction intervals constructed from the test RMSE. The projection captured the expected seasonal cycle, with summer values peaking around 24 to 25 degrees and winter minima between 22 and 23 degrees, and all forecasted temperatures remained above the cold wave threshold. But the authors are emphatic that this absence of forecasted cold waves must not be read as a prediction of no cold wave risk. A model that misses roughly 78 percent of historical cold days would likely miss actual cold waves in 2027 as well. The model was trained on data ending in 2011 and cannot account for climate regime shifts or changing winter patterns since then, and the winter prediction intervals, spanning roughly plus or minus 2 to 3 degrees, are wide enough that a downward fluctuation could still push temperatures below 10 degrees. The researchers stress that disaster management agencies, health authorities and local governments should continue normal winter preparedness measures from November through February regardless of this exploratory projection, and that operational decisions should rely on routine seasonal forecasts from national meteorological services.

The study’s implications reach beyond Bangladesh. The finding that deep learning substantially outperforms linear statistical models for daily temperature echoes results from across South Asia, including ARIMA-based temperature analysis in Karachi, Pakistan, and an STL-ARIMA-LSTM hybrid for heatwave forecasting in Rajshahi that achieved an RMSE of 1.18 degrees. The low recall for rare events is also not unique to this work; studies of heatwave and flood classification have reported similar struggles with imbalanced datasets, where high overall accuracy conceals poor detection of the very events that matter most. The authors recommend that future operational systems adjust the classification threshold to balance precision and recall, incorporate perceived temperature metrics such as the wind chill index, and integrate humidity, wind speed, cloud cover and large-scale climate indices like ENSO, all of which were absent from the current univariate framework.

The researchers outline a clear roadmap for strengthening the approach. Expanding the analysis to multiple stations across Bangladesh would test spatial transferability and reveal regional patterns in cold wave occurrence. Advanced techniques for handling class imbalance, including synthetic minority oversampling, focal loss and cost-sensitive learning, could raise recall at some cost to precision. Linking temperature forecasts to health outcome data such as cold-related mortality and hospitalization rates would allow health-relevant alert thresholds to be defined and would provide direct validation of the model’s usefulness as an early warning tool. Probabilistic methods such as quantile regression forests and Bayesian neural networks could extend forecast horizons while quantifying uncertainty more honestly. For now, the study positions the LSTM model as a powerful temperature forecasting instrument rather than a standalone cold wave alarm, a distinction that could shape how machine learning is deployed to protect vulnerable communities across the region.

Subject of Research: Machine learning forecasting of daily minimum temperatures and cold wave events in Mymensingh district, Bangladesh

Article Title: Forecasting cold wave in Bangladesh: a validated machine learning approach for early warning and vulnerability reduction

Article References: Akther, S., Hussain Khan, M. M., Chowdhury, S., Das, A., Rahman, M., & Rois, R. (2026). Forecasting cold wave in Bangladesh: a validated machine learning approach for early warning and vulnerability reduction. BMC Environmental Science, 3(1), Article 17. https://doi.org/10.1186/s44329-026-00058-6

Image Credits: AI Generated

DOI: 10.1186/s44329-026-00058-6

Keywords: cold wave, Bangladesh, LSTM, machine learning, temperature forecasting, early warning, Mymensingh, ARIMA, hybrid models, time series, climate extremes, public health