The appeal of long-term traffic forecasting is easy to understand. Short-term predictions, covering the next few minutes or an hour, help with immediate tasks such as adjusting signal timings or warning drivers of congestion ahead. But the decisions that really shape a city, from planning road maintenance windows to scheduling public transport fleets and managing emergency response capacity, depend on knowing what traffic will look like many hours or even days from now. At those horizons, errors compound quickly. Small mistakes in how a model interprets today’s conditions cascade into large mistakes about tomorrow’s rush hour, which is why most operational systems have stayed firmly in the short-term regime.

The researchers’ answer is an architecture that splits the problem into two complementary parts. The convolutional neural network, or CNN, acts as the spatial analyst. CNNs, which rose to fame in image recognition, excel at detecting local patterns in grid-like data, and multi-sensor traffic readings can be arranged so that the network learns how congestion at one intersection relates to conditions on neighboring streets. The CNN component extracts these spatial features from the streams of data arriving from IoT devices deployed across the transportation network, effectively giving the model a snapshot of the road network’s geometry of congestion at each moment.

The second half of the hybrid is the bidirectional long short-term memory network, or BiLSTM. LSTM networks are a form of recurrent neural network designed to remember information over long sequences, using gated cells that decide what to store, what to discard, and what to pass forward. This makes them well suited to temporal dependencies in traffic, such as the daily rhythm of morning and evening peaks, the weekly cycle that distinguishes weekdays from weekends, and the slow drift of seasonal patterns. The bidirectional twist means the network processes the sequence both forward and backward, so its understanding of any given time step is informed by what came before and what came after, allowing it to capture context that a one-directional reader would miss.

Chained together, the two components form a pipeline that mirrors the structure of the problem itself. Raw sensor readings flow into the CNN, which compresses the spatial state of the network into informative feature representations. Those features then feed into the BiLSTM, which models how they evolve through time and projects them forward across the forecast horizon. This division of labor is the key to the model’s stability: rather than forcing a single network to learn both where traffic happens and how it changes, the hybrid assigns each task to the architecture best built for it.

To test the approach, the team evaluated the model on traffic datasets representing four major locations in Cairo: Misr El-Gedida, Nasr City, Downtown, and New Cairo. Cairo is a demanding proving ground for any forecasting system. Its streets carry enormous and highly variable loads, its traffic patterns are shaped by a dense mix of private cars, informal transport, and commercial vehicles, and disruptions propagate through the network in ways that are difficult to anticipate. A model that performs well on Cairo data has faced some of the messiest conditions urban traffic can offer.

The results, measured using standard error metrics including mean absolute error and root mean square error alongside overall prediction accuracy, showed that the hybrid CNN–BiLSTM model significantly outperformed both traditional machine learning approaches and single deep learning models. That comparison matters because it isolates the value of the hybrid design. Classical methods such as autoregressive statistical models and simpler neural networks have historically struggled to represent the intertwined spatial and temporal structure of traffic, while single deep architectures tend to sacrifice one dimension for the other. The hybrid’s advantage across all three evaluation metrics suggests that the CNN and BiLSTM components are genuinely complementary rather than redundant.

Perhaps the most striking finding is the model’s endurance. It maintained stable performance for prediction horizons up to 72 hours, meaning city operators could in principle run the model on a Monday and receive a usable picture of Thursday’s traffic. The authors highlight this stability as the feature that makes the system practical for real-world smart city traffic management, where forecasts need to be reliable enough to act on, not merely interesting in a laboratory setting. Stable multi-day forecasts could support proactive signal control, dynamic lane management, better-informed logistics routing, and earlier warnings for planned events that will stress the network.

The study also sits within a broader wave of research into deep learning for transportation. Recent years have seen graph neural networks, attention-based architectures, and transformer models all applied to traffic prediction, each trying to capture the spatial-temporal character of urban mobility in different ways. The CNN–BiLSTM approach represents a pragmatic strand of this effort: it builds on two of the most mature and well-understood deep learning building blocks, which makes the resulting system easier to train, deploy, and maintain than more exotic architectures. The authors have made their source code and experimental resources publicly available on GitHub, a step that should help other researchers reproduce the results and adapt the model to their own cities.

For the smart cities movement, the work offers a concrete demonstration of what IoT infrastructure can deliver when paired with the right analytical machinery. Cities have spent the past decade blanketing their streets with sensors, cameras, and connected devices, generating torrents of data that often far exceed what existing systems can exploit. Models like the one developed in Cairo are a step toward converting that raw data deluge into operational foresight. If the hybrid approach generalizes beyond Cairo’s challenging roads, the practical implications could reach far beyond traffic lights and congestion maps, informing everything from urban planning decisions to emissions reduction strategies that depend on anticipating where and when vehicles will cluster. The three-day traffic forecast, once a distant ambition, is now a demonstrated capability, and cities around the world will be watching how far it can be pushed.

Subject of Research: Long-term urban traffic flow forecasting using a hybrid CNN–BiLSTM deep learning model based on IoT sensor data

Article Title: Smart cities long-term traffic flow forecasting using a hybrid CNN–BiLSTM model based on IoT

Article References: Hussien, M. A. E., & AboRizka, M. (2026). Smart cities long-term traffic flow forecasting using a hybrid CNN–BiLSTM model based on IoT. Neural Computing and Applications, 38(19), Article 789. https://doi.org/10.1007/s00521-026-12423-2

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12423-2

Keywords: traffic forecasting, CNN, BiLSTM, deep learning, smart cities, Internet of Things, IoT sensors, long-term prediction, Cairo, intelligent transportation, machine learning, urban mobility

The core of the research is a framework the authors call ICO-2DQN, a hybrid of an Intelligent Cuckoo Optimization algorithm and a Dueling Deep Q-Network. The system was trained on 4,000 rows of biomechanical data recorded from martial artists performing fundamental techniques—chopping, kicking, and grappling—while wearing IoT sensors. Each record captured a single movement instance, with sensor-derived features describing joint motion, force distribution, body stability, and movement performance. The goal was ambitious: build a machine that could not merely classify a movement as good or bad, but generate individualized corrective feedback the way an expert coach would, only faster and with quantitative precision.

Before any learning could happen, the raw sensor data required careful preparation. The researchers handled missing values to prevent biased biomechanical estimates, then applied min–max normalization to scale every heterogeneous feature onto a common range from zero to one. This step matters more than it might appear: joint angles measured in degrees, forces measured in newtons, and velocities measured in radians per second live on wildly different scales, and without normalization a single large-magnitude feature could dominate the learning process. The preprocessing ensured that every biomechanical signal contributed equally to the model’s understanding of a technique.

The next stage used Independent Component Analysis, a statistical technique that decomposes complex, mixed sensor signals into their underlying independent sources. Wearable sensors pick up a tangle of overlapping information—muscle-driven forces, gravitational effects, sensor noise, and the compound motion of linked body segments all arrive blended together. ICA separates that mixture, isolating distinct biomechanical patterns that would otherwise remain hidden. The researchers retained twelve independent components, a number chosen to reduce redundancy while preserving the essential structure of each movement. Compared with alternatives such as kernel ICA or autoencoders, linear ICA offered faster execution and lower computational cost, an important consideration for a system intended to run alongside real-time IoT coaching.

From the cleaned and decomposed signals, the system derived the biomechanical quantities that define martial arts technique: joint angles, angular velocities, center-of-mass stability, and force delivery. These four quantities form the state space of the reinforcement learning environment. At every time step, the learning agent observes a vector containing the practitioner’s joint angles, angular velocities, balance stability, and force distribution. Its action space consists of the corrective instructions a coach might issue—posture correction, balance adjustment, movement speed modification, or simply continuing the current motion. A weighted reward function scores each executed movement based on posture accuracy, balance stability, and movement similarity, so the agent learns to issue feedback that maximizes the quality of the student’s technique.

The Dueling Deep Q-Network at the heart of the system introduces a clever architectural refinement over standard deep Q-learning. Rather than estimating the value of every state-action pair directly, the dueling architecture decomposes the action-value function into two separate streams: a state-value function that captures how good a biomechanical position is in itself, and an advantage function that captures how much a particular action improves on the average. This decoupling is well suited to martial arts, where many states have similar values regardless of the action taken, and it accelerates convergence in the high-dimensional, continuous movement spaces that full-body biomechanics produce. The network uses convolutional layers for feature extraction before splitting into the two value streams, which are then recombined into Q-values that guide the selection of optimal training strategies.

Reinforcement learning systems, however, are notoriously sensitive to their hyperparameters—learning rates, discount factors, and the weights inside the reward function. Poorly tuned, an agent converges slowly or oscillates between policies. That is where the cuckoo enters. The Intelligent Cuckoo Optimization algorithm draws its inspiration from the brood parasitism of real cuckoos, which lay their eggs in the nests of host birds. In the algorithmic version, candidate solutions are cuckoo eggs, each habitat receives a variable number of eggs within defined bounds, and an egg-laying radius determines how far solutions scatter in the search space. The algorithm also incorporates Lévy flight behavior, the characteristic pattern of long jumps interspersed with short steps that cuckoos use when seeking new habitats. Applied here, the ICO dynamically tunes the 2DQN’s hyperparameters and reward weightings, replacing weak solutions with fitter ones and steering the learning process away from local optima.

The experimental results are striking. Implemented in Python and evaluated with five-fold cross-validation on the IoT sensor dataset, the ICO-2DQN model achieved an accuracy of 98.7 percent, a precision of 98 percent, a recall of 96.5 percent, and an F1-score of 97.1 percent, with a root-mean-square error of 0.15. In real-time testing, the system recorded an accuracy of 98.2 percent, an average F1-score of 0.981, and a feedback delay of just 82 milliseconds—fast enough for a student to receive a correction while the movement is still fresh. Statistical validation with paired t-tests and 95 percent confidence intervals confirmed the gains were significant at p < 0.001. During training, the model’s cumulative reward climbed from roughly 20 to 98 while training loss fell from about 1.12 to 0.04, and the exploration rate decayed from 1.0 to 0.1 as the agent shifted from exploration to exploitation.

The comparison against baseline models reinforces the case. When retrained and tested on the same dataset under identical conditions, competing frameworks—including two-stream CNNs, a 3D-CNN combined with LSTM, a Vision Transformer paired with a DQN, a Sunflower Optimization multi-column CNN, and classical support vector machines—fell short on accuracy, error rates, feedback latency, and stability. Each baseline carried known weaknesses: optical-flow dependence and computational cost in two-stream CNNs, limited interpretability in 3D-CNN plus LSTM models, heavy data and compute demands in ViT-DQN, and motion-blur sensitivity in YOLO plus LSTM approaches. The dueling architecture’s explicit modeling of sequential action-state transitions, such as the shift from a kick to a grappling exchange, proved better suited to the spatiotemporal dependencies of martial arts than any of the alternatives.

Evaluation across skill levels showed the framework adapting its behavior for beginner, intermediate, and advanced practitioners while maintaining high movement recognition accuracy and reliable biomechanical assessment throughout. The authors are candid about the remaining hurdles: the system depends on accurate, fine-grained motion data, and inconsistent or faulty sensor readings could corrupt the analysis; real-time deployment also demands robust wearable hardware that resource-limited training environments may lack. Still, the trajectory is clear. The researchers envision future systems that adapt to specific martial arts styles, body types, and skill levels, integrating live motion capture to adjust training programs automatically with an emphasis on both performance and injury prevention. Biomechanical monitoring can already flag irregular loading patterns and fatigue-related deviations before they become injuries, and a reinforcement learning agent that prescribes corrective exercises or modulates training intensity in advance could make practice substantially safer. If the approach matures, the centuries-old model of learning by imitation may gain a tireless digital partner—one that sees what no coach can, and answers in milliseconds.

Subject of Research: Deep reinforcement learning combined with biomechanical movement analysis for adaptive martial arts training

Article Title: Innovative martial arts teaching methods based on deep reinforcement learning and biomechanics of movement

Article References: Zheng, L., & Zhang, J. (2026). Innovative martial arts teaching methods based on deep reinforcement learning and biomechanics of movement. Discover Artificial Intelligence, 6(1), Article 1307. https://doi.org/10.1007/s44163-026-01887-9

Image Credits: AI Generated

DOI: 10.1007/s44163-026-01887-9

Keywords: martial arts, deep reinforcement learning, biomechanics, Dueling DQN, cuckoo optimization, IoT sensors, Independent Component Analysis, sports training, movement analysis, personalized coaching, injury prevention, machine learning

The research, led by Freddy Artadima Silaban of Institut Teknologi Bandung and colleagues, was conducted at an experimental site in Tajur, Bogor, Indonesia. Over an 81-day shallot growing season from October to December 2025, an Internet of Things sensor network recorded ten variables across 300 polybags: soil moisture, soil temperature, soil pH, electrical conductivity, nitrogen, phosphorus, and potassium levels, along with light intensity, air temperature, and air humidity. The resulting datasets were substantial, comprising 269,080 observations for nitrogen, 286,232 for phosphorus, and 338,242 for potassium. Each observation was labeled low, medium, or high according to the nutrient treatment levels applied in the experimental design, with nitrogen dosages ranging from 0.7 to 2.8 grams, phosphorus from 0.8 to 3.2 grams, and potassium from 1.0 to 3.0 grams.

The team evaluated three neural architectures suited to tabular data: a conventional multilayer perceptron, TabTransformer, which uses attention mechanisms to learn contextual relationships between features, and TabNet, which applies sequential attention for adaptive feature selection. The TabTransformer teacher model, with roughly 830,000 parameters, was the largest of the three, while the MLP teacher contained about 69,000 parameters and TabNet about 116,000. Rather than relying on a single training run, the researchers trained every model across five fixed random seeds and evaluated results on a chronological data split, in which the first 70 percent of dates were used for training, the next 15 percent for validation, and the remainder for testing. This temporal split prevents information from the future leaking into training, a critical safeguard when working with time-stamped sensor data.

The heart of the study is knowledge distillation, a compression technique in which a compact student model is trained to mimic the output distribution of a larger teacher model. The researchers froze each teacher’s parameters and trained a smaller student from the same architectural family using a combined loss function: 60 percent weight on standard cross-entropy against the true labels and 40 percent weight on Kullback-Leibler divergence against the teacher’s softened probability distribution, generated with a temperature of 4.0. The softening spreads the teacher’s confidence across classes, revealing not just what the teacher predicted but how certain it was about alternatives, information that helps the smaller student generalize. The compression results were dramatic. The MLP student shrank to just 1,795 parameters, a 38.65-fold reduction; the TabTransformer student dropped to 109,891 parameters, a 7.55-fold reduction; and the TabNet student reached 10,484 parameters, an 11.04-fold reduction.

The accuracy story, however, was more nuanced. For the nitrogen dataset, the TabNet teacher achieved the highest average accuracy at 91.26 percent, while the distilled TabTransformer student reached 84.99 percent. On the potassium dataset, the distilled TabTransformer student actually outperformed its teacher, achieving 54.55 percent accuracy. But in other configurations, distillation offered no clear benefit: for phosphorus, the distilled MLP and TabNet students fell 5.04 and 5.36 percentage points below their teachers, with confidence intervals lying entirely below zero. The authors are candid about this variability, noting that knowledge distillation does not establish a uniform pattern of improvement across all dataset and architecture combinations. What the technique reliably delivers is a far smaller model, and in several cases the distilled students matched or exceeded both their teachers and control students trained from scratch with identical architectures.

A particularly revealing part of the study is the feature ablation analysis, which probed what the models were actually learning. The researchers systematically removed individual features and measured the impact on accuracy. Surprisingly, removing the target nutrient channel itself, such as nitrogen readings from the nitrogen dataset, produced almost no change in performance. Even removing all three NPK channels together changed accuracy only slightly. Accuracy collapsed only when the models were restricted to nutrient and electrical conductivity inputs alone, with drops of roughly 14 to 32 percentage points. This demonstrates that the classification signal is distributed across the full combination of soil and environmental variables rather than residing in any single sensor channel. The authors caution that the low, medium, and high labels represent experimental treatment levels, not laboratory-verified nutrient availability, so the models classify treatment status rather than replacing chemical soil analysis.

Before deployment, the team verified that the conversion pipeline preserved predictions exactly. Models were exported from PyTorch to the Open Neural Network Exchange format and then compiled into TensorRT engines on the target hardware in three numerical precision configurations: FP32, FP16, and INT8. All 27 engine combinations, spanning three datasets, three architectures, and three precision levels, achieved 100 percent prediction agreement with the original models, with maximum logit errors in the range of one millionth to nearly one ten-thousandth. INT8 calibration used 4,096 samples drawn deterministically from training data only, keeping test data untouched. On the test set, INT8 precision maintained identical accuracy and macro-F1 scores to FP32 in every configuration, while FP16 shifted accuracy by at most 0.01 percentage points.

The edge inference benchmarks on an NVIDIA Jetson Orin NX, running in its 15-watt power mode, revealed striking differences between architectures. The distilled MLP was the clear efficiency champion, with mean latencies between 0.128 and 0.262 milliseconds and throughputs reaching nearly 7,600 samples per second, consuming roughly 1 millijoule per inference. The TabTransformer and TabNet students required roughly 4 to 6 millijoules per inference with latencies around 0.7 to 2 milliseconds. Notably, lower numerical precision did not always translate to faster inference: FP16 consistently reduced tail latency for TabTransformer, cutting the p99 latency for the nitrogen task from 8.5 milliseconds to 2.7 milliseconds, but INT8 sometimes increased latency for TabNet, reaching nearly 10.2 milliseconds on the potassium task. All 27 configurations passed a 300-second continuous inference stability test without failures or prediction mismatches, with GPU temperatures holding steady between 61 and 64 degrees Celsius.

The study also benchmarked the neural models against classical gradient-boosting baselines: XGBoost, LightGBM, and CatBoost, evaluated with the same seeds and temporal split. For nitrogen, LightGBM and CatBoost remained highly competitive, with CatBoost reaching 90.34 percent accuracy. For potassium, however, the distilled TabTransformer student outperformed all classical baselines on accuracy and macro-F1, though CatBoost achieved higher balanced accuracy. On the edge device, LightGBM and CatBoost achieved sub-millisecond p99 latencies on a single CPU thread, while XGBoost was considerably slower and more energy-hungry. The authors emphasize that the neural models ran on TensorRT GPU acceleration while the baselines used native CPU execution, so the deployment comparison reflects measured performance on the same hardware rather than identical computing backends.

The broader significance of this work lies in its honest, multi-dimensional evaluation framework. Rather than chasing a single accuracy figure, the researchers simultaneously assessed predictive robustness across seeds, compression ratios, feature dependencies, conversion fidelity, numerical precision, latency, throughput, energy consumption, and stability. Their conclusion is measured: knowledge distillation reliably shrinks tabular neural models for edge deployment, but its accuracy benefits depend on the interplay of dataset and architecture, and classical baselines remain formidable competitors that should never be dismissed. The team acknowledges limitations, including the absence of polybag identifiers for grouped evaluation, the lack of laboratory validation of nutrient labels, and the brevity of the 300-second stability tests. Future work, they suggest, should partition data by physical growing unit, validate labels through laboratory analysis, and conduct long-term field trials measuring thermal behavior and energy consumption under real operating conditions. For precision agriculture, the message is clear: the era of running meaningful AI directly in the field, without the cloud, is arriving one compressed model at a time.

Subject of Research: Knowledge distillation and edge deployment of tabular neural networks for classifying soil nutrient availability in shallot cultivation

Article Title: Knowledge distillation and edge deployment of tabular neural models for soil nutrient availability status classification

Article References: Silaban, F. A., Trilaksono, B. R., Mahayana, D., Maharijaya, A., Gunawan, E., Rosliani, R., & Prathama, M. (2026). Knowledge distillation and edge deployment of tabular neural models for soil nutrient availability status classification. Smart Agricultural Technology, 15, Article 102580. https://doi.org/10.1016/j.atech.2026.102580

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102580

Keywords: knowledge distillation, edge AI, precision agriculture, soil nutrients, TabTransformer, TabNet, NVIDIA Jetson Orin NX, TensorRT, IoT sensors, model compression, shallot cultivation, tabular deep learning

The commercial stakes are considerable. The global fresh strawberry market has been estimated at roughly 20.7 billion US dollars in 2024, with projections exceeding 30 billion dollars by the mid-2030s as year-round demand grows. In the United Kingdom alone, official statistics value the strawberry crop at 389 million pounds for 2024. At daily resolution, even modest forecasting errors ripple through the entire operation: harvest crews are scheduled against predicted volumes, packing lines and storage capacity are booked in advance, and supermarkets hold contracts that assume reliable supply. A model that misses a peak picking day by a wide margin creates labour inefficiencies on one hand and post-harvest waste on the other.

What makes strawberries so hard to forecast is that daily yield is not a string of independent numbers but a biologically structured time series. Production follows a recognisable lifecycle: output starts at zero before harvest begins, climbs as plants enter peak fruiting, and falls back to zero as the season ends. That rise-peak-decline shape is governed by cultivar, planting density, accumulated temperature and humidity, and the decisions of farm managers. The researchers argue that a realistic forecasting framework must satisfy three conditions at once. It must preserve the natural lifecycle, including genuine zero-yield periods at season boundaries. It must enforce strict temporal alignment so that only information available before harvest can be used as a predictor. And it must test models on unseen future seasons rather than interpolating within the data it has already seen.

That last condition turns out to be the study’s sharpest methodological point. In agricultural machine learning, it is dangerously easy to inflate performance through information leakage, for example by letting weather recorded after a harvest event act as a predictor, or by using random train-test splits on strongly seasonal data. The team deliberately excluded aggregate productivity indicators such as yield per hectare and percentage picked from the feature set, because those variables indirectly encode the answer. Environmental sensor readings from the farm’s SoilMoistureSense platform, covering temperature and relative humidity, were aggregated strictly within calendar-day boundaries and merged by date so that no future information could slip into the pipeline.

The dataset itself is a portrait of messy commercial reality rather than a tidy laboratory experiment. It was assembled from three sources: annual weekly planning spreadsheets, 189 individual daily forecast files recording both the farm’s own operational predictions and actual harvested weights, and sub-daily environmental sensor exports. After cleaning and harmonisation across years, the modelling dataset contained 2,280 daily observations, which rose to 3,540 after the researchers appended seven zero-yield days at the start and end of each harvest cycle. That padding step, an ablation analysis showed, mattered a great deal: removing it degraded the best model’s mean absolute error from about 340 kilograms to nearly 611 kilograms, while extending padding to 14 days lowered absolute error further but reduced explained variance. The seven-day configuration offered the best balance.

Feature engineering was deliberately biological. Rolling averages of prior yield over 1, 3, 7, 14 and 21 days captured what the authors call production memory, the tendency of strawberry harvests to persist across neighbouring days as overlapping cohorts of fruit mature and ripen. Rolling seven-day temperature averages and 14-day humidity averages represented accumulated environmental exposure, reflecting how heat and moisture stress modulate fruit development. Plant counts from the planning spreadsheets served as a static capacity indicator for each field. Together these features let the models learn whether production is rising, plateauing or declining, without ever violating temporal causality.

The model comparison spanned thirteen approaches, from simple ridge regression through gradient boosting frameworks such as XGBoost, LightGBM and CatBoost, to compact neural networks including a multilayer perceptron, a small LSTM and a temporal convolutional network, and finally TabPFN, a pretrained tabular foundation model. The results were striking. Under a conventional random 80/20 split, TabPFN achieved the lowest error with a mean absolute error of 340.18 kilograms and an R-squared of 0.733, while the neural sequence models performed worse than a constant mean predictor, producing negative R-squared values. Leave-one-out cross-validation told the same story, with TabPFN again leading at a mean absolute error of 338.09 kilograms. The authors note that with only six seasons of data, the strong inductive biases of tree ensembles and pretrained tabular models matter far more than architectural depth, which is precisely why the from-scratch deep learning baselines struggled.

The most deployment-relevant test, however, was forward-year validation: train on 2020 through 2024, then predict the entirely unseen 2025 season. Under this protocol TabPFN remained the strongest performer with a mean absolute error of 438.53 kilograms and an R-squared of 0.696, followed by Random Forest, Histogram Gradient Boosting and Extra Trees. The rise in error relative to the random split is not a failure, the researchers argue, but an honest measure of genuine interannual variation in weather, phenology and management. Permutation importance analysis on the 2025 test season confirmed that the models rely on biologically sensible variables: recent yield lags dominate the ranking, and interaction features combining recent production with temperature and humidity contribute measurably, suggesting that weather modulates an established production trajectory rather than acting as an independent driver.

The study is candid about its limits. All data come from a single farm, so the work demonstrates temporal generalisation to an unseen season within one commercial setting, not transfer across farms or regions. The available sensors captured only temperature and humidity, and the authors identify the absence of solar radiation, day length, irrigation records and growing degree days as a likely explanation for a recurring error mode: the systematic underestimation of peak harvest days, when similar recent conditions can conceal very different yield potential. Adding phenology-aware signals such as flower or truss counts, whether recorded manually or by computer vision, is flagged as a plausible route to better peak capture.

Even so, the implications are significant for an industry built on thin margins and perishable goods. The findings suggest that daily yield forecasting can be integrated into commercial decision-support systems using the data farms already collect, without exotic sensors or vast historical archives. In small, discontinuous seasonal datasets, the study concludes, strong inductive bias and regularisation beat model complexity, and rigorous forward-year evaluation is the only honest yardstick of what a forecasting system will actually deliver when the next season begins. For growers weighing the promise of agricultural AI, that combination of realism and rigour may be the most valuable harvest of all.

Subject of Research: Daily strawberry yield forecasting with artificial intelligence in commercial horticulture

Article Title: Daily strawberry yield forecasting using artificial intelligence in commercial horticulture

Article References: Ahmed, S., & Caldwell, N. H. (2026). Daily strawberry yield forecasting using artificial intelligence in commercial horticulture. Smart Agricultural Technology, 15, Article 102583. https://doi.org/10.1016/j.atech.2026.102583

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102583

Keywords: strawberry, yield forecasting, machine learning, TabPFN, deep learning, precision agriculture, time series, commercial horticulture, feature engineering, forward-year validation, IoT sensors, UK farming

A team of researchers led by Kashif Khaqan of Federation University Australia, working with colleagues at Adelaide University and partners in industry, has published one of the most comprehensive maps to date of the scientific literature on digital twins in agriculture. Their bibliometric analysis, published in the journal Discover Agriculture, examined 657 Scopus-indexed publications spanning 23 academic disciplines and 87 countries. The headline finding is a striking surge of activity: more than four-fifths of all relevant documents identified in the dataset were published in just the last three years of the study period, with 2025 alone contributing 216 papers, or nearly 33 percent of the total. What was once a niche curiosity has become one of the most dynamic research frontiers in agricultural science.

The study defines a digital twin as a closed-loop system in which real-time data, decision-making, and automated actions continuously interact with and influence a physical asset. This is what separates a true twin from an ordinary computer model. A crop simulation can be run once and left alone; a digital twin is fed constantly by Internet of Things sensors measuring soil properties, microclimates, crop status, and equipment performance, and its outputs flow back into the physical world as irrigation commands, machinery adjustments, or management alerts. The researchers argue that this bidirectional connection gives digital twins a level of fidelity and usefulness that traditional modelling and simulation approaches cannot match, effectively turning a farm into a system that can be rehearsed, predicted, and optimised before decisions are executed in the field.

The numbers behind the analysis reveal where the intellectual centre of gravity currently sits. China leads global output with 118 publications, representing 17.9 percent of the total and attracting 1,129 citations, followed by India with 105 papers and the United States with 66. The United States ranks second in citations with 1,209, while Australia, despite a smaller share of publications, has accumulated the highest total citations at 1,371, suggesting outsized influence per paper. On the institutional stage, Wageningen University & Research in the Netherlands stands out as the single most productive organisation with 21 publications, ahead of China’s Ministry of Agriculture and China Agricultural University. The top five countries together produced more than half of the entire corpus, and 45 of the 87 publishing nations met a threshold of at least five shared collaborations, indicating a genuinely international but unevenly distributed research network.

Perhaps the most revealing part of the study is its keyword co-occurrence analysis, which distilled 4,726 author keywords down to 105 that appeared at least ten times, and then clustered them into five interconnected themes. The first cluster captures the technological backbone: artificial intelligence, big data, cyber-physical systems, real-time monitoring, decision support systems, network security, and predictive maintenance. The second focuses on digital crop management, where machine learning, deep learning, and reinforcement learning allow twins to improve their predictions of crop growth, irrigation needs, and environmental conditions over time, with particular strength in controlled environments such as greenhouses. The third cluster highlights sensing and autonomy: drones, remote sensing, 3D reconstruction, agricultural robots, edge computing, and blockchain. The fourth links the technology to climate-smart agriculture, sustainability, life cycle assessment, food security, and climate change adaptation. The fifth centres on the IoT-driven ecosystem of sensors and connected devices that underpins agriculture 4.0.

Dominance by computer science is both the engine and the weakness of the field. Computer Science accounts for 62.2 percent of publications and Engineering for 48.5 percent, while Agricultural and Biological Sciences contributes just 23.7 percent. The authors read this as a disciplinary imbalance: the machinery of digital twins is being built faster than agronomists and practitioners can apply it. Journal patterns tell a similar story. Computers and Electronics in Agriculture leads the field with 21 articles and 867 citations and a journal impact factor of 8.9, while broader outlets such as Agriculture and Applied Sciences fill out the top tier. Conference papers, at 36 percent of the corpus, slightly outnumber research articles at 32 percent, a signature of a young field still working out its methods in real time.

The gaps are as informative as the clusters. Keywords associated with disease management, climate resilience, agricultural supply chains, and advanced decision-support systems appeared less frequently and linked weakly to the core themes. In other words, most existing digital twins are good at watching and predicting, but poor at deciding. The analysis identifies decision support and disease management as the key research deficits, and the authors call for decision-oriented frameworks that close the loop from sensor data through simulation to a concrete, defensible recommendation for the farmer. On the research side, they argue that twins enable scenario testing without risking crop loss, allow controlled experiments on plant-soil-microbe interactions and pest propagation, and could eventually be aggregated into regional and global networks for supply chain optimisation and food security planning, provided that standardised protocols for data collection, integration, and validation are established.

For practice, the pathway to adoption runs through high-value and controlled environments first. The study suggests farmers and agribusinesses begin with greenhouses, orchards, and farms that already have sensor infrastructure, where twins can deliver measurable gains in precision irrigation, nutrient management, yield forecasting, and input-use efficiency. Integration with existing farm management systems, autonomous machinery, and supply chain processes extends the benefits toward traceability and sustainability. But the barriers are real: high implementation costs, technological complexity, a shortage of expertise, interoperability problems between hardware and software platforms, and unresolved questions of data security and privacy. The authors point to pilot deployments, cost-effective sensor networks, user-friendly interfaces, and training programmes as the practical instruments for widening adoption beyond large commercial operations to smallholders.

Policy emerges as the decisive variable. Because digital twins depend on real-time data, sensor networks, and robust connectivity, countries with strong national innovation strategies and Industry 4.0 investment, such as China, dominate output, while many smallholders in lower-income regions cannot afford the underlying infrastructure. The study recommends subsidies, grants, low-interest financing, and shared equipment schemes to prevent diffusion from being confined to industrial-scale farming, alongside interoperability standards developed with industry bodies and universities, and clear regulations on data privacy and ownership to build farmer trust. The analysis has limitations worth noting: it relies on a single database, English-dominated literature, and a deliberately narrow search string pairing digital twin with agriculture, which may exclude adjacent work using terms like smart farming or cyber-physical systems. Yet the trajectory is unmistakable. In a world of tightening water, unstable weather, and rising food demand, the ability to rehearse a bad harvest virtually before it happens may prove one of agriculture’s most consequential tools, and the research community is clearly betting that the twin will soon be as fundamental to a farm as the soil beneath it.

Subject of Research: Bibliometric analysis of digital twin applications in agriculture

Article Title: Exploring digital twins in agriculture with current applications and implications for research, policy and practice

Article References: Khaqan, K., Gautam, P., Paudel, B., Liang, S., Van Duc Long, N., Parvin, M. I., Munigoti, K., Sasanelli, N., Hessel, V., & Sandhu, H. (2026). Exploring digital twins in agriculture with current applications and implications for research, policy and practice. Discover Agriculture, 4(1), Article 278. https://doi.org/10.1007/s44279-026-00760-8

Image Credits: AI Generated

DOI: 10.1007/s44279-026-00760-8

Keywords: digital twins, agriculture, bibliometric analysis, smart farming, Internet of Things, artificial intelligence, precision agriculture, climate-smart agriculture, machine learning, drones, IoT sensors, agricultural policy

The study, led by Hanjie Xu of Zhongguancun Lab and Beihang University together with Ju Ren of Tsinghua University, Jie Liu of Beijing Information Science and Technology University, Zhenya Ma, Wang Huang of Central South University, and Jianwei Niu of Beihang University, was published in the International Journal of Machine Learning and Cybernetics. The team’s model, which fuses wavelet analysis, graph neural networks, and the newly popular Mamba sequence architecture, outperformed a wide range of established baselines in extensive benchmark testing. The work addresses one of the most persistent challenges in applied machine learning: detecting anomalies in multivariate time series, where dozens or hundreds of correlated sensor channels evolve together and a fault in one can ripple through the others in complex ways.

Each of the three components of the new model plays a distinct role. Wavelet representations, the first ingredient, give the network a way to see a signal at multiple resolutions simultaneously. Traditional Fourier analysis tells an engineer which frequencies exist in a signal but says little about when they occurred, a limitation known as the time frequency trade off. Wavelet transforms solve this by decomposing a signal into localized oscillations of different scales, so the model can capture both a slow drift lasting minutes and a sharp transient lasting milliseconds. By enhancing the raw inputs with these wavelet features, the architecture extracts far richer representations of each moment in the stream than models that read the raw sensor values alone.

The second ingredient, graph neural networks, addresses the fact that sensor data is rarely a set of independent channels. In a water treatment plant, a pressure sensor upstream is physically and causally linked to a flow sensor downstream; in a server cluster, the temperature readings of adjacent racks move in concert. Graph neural networks encode these relationships explicitly, treating each sensor as a node and the learned or measured dependencies as edges. Message passing across the graph allows the model to build a picture of how the whole system should behave as a coherent network, and to flag moments when that coherence breaks down. This topology learning capability, the authors argue, is essential for detecting anomalies that only become visible when several channels are viewed together.

The third component, Mamba, is one of the most talked about architectures in artificial intelligence today. Introduced by Albert Gu and Tri Dao in late 2023, Mamba is a selective state space model that processes sequences in linear time, in contrast to the quadratic cost of transformer attention. For time series workloads, where context windows can stretch across thousands of timesteps and the model must run on resource constrained edge devices, that efficiency matters enormously. Mamba’s selective mechanism lets the network decide dynamically which parts of the past to remember and which to forget, giving it a powerful ability to follow long range sequential patterns without the memory burden that attends older recurrent designs.

What makes the new model distinctive is the way these three ideas operate together rather than in sequence. The wavelet enhanced representations give the sequence model a frequency aware view of each channel, the graph neural network binds those views into a coherent picture of system wide behavior, and the Mamba backbone explores how the entire enriched representation evolves through time. In effect, the model asks three questions at once: how is each sensor oscillating, how are the sensors related, and how are those patterns changing? An anomaly is registered when any of these three questions produces an answer that deviates from what the model has learned to expect from normal operation.

The evaluation was extensive. The researchers tested their architecture against a broad collection of time series anomaly detection baselines, spanning the main families of methods that have dominated the field over the past decade. Those baselines include long short term memory networks, which have been used for everything from spacecraft telemetry to rail transit monitoring; variational autoencoders and generative adversarial networks that learn to reconstruct normal data and flag what they cannot reconstruct; transformer models such as TranAD and the Anomaly Transformer, which use attention to model dependencies across time; and earlier graph based detectors, including graph attention networks and graph deviation networks, which pioneered the topological view of multivariate data. Across this competitive landscape, the wavelet enhanced Mamba and graph network model outperformed many prevailing approaches.

The significance of the result extends beyond leaderboard positions. Anomaly detection sits at the foundation of industrial safety, cybersecurity, financial monitoring, and healthcare. Earlier studies in the literature the team cites show detectors deployed in industrial IoT systems, smart city traffic networks, financial markets analyzed with principal component analysis and neural networks, and clinical settings where deep learning reviews promise earlier warnings. Each of these domains has its own constraints: latency requirements that forbid heavy computation, noisy training sets, and distribution shifts that arise as machines age or seasons change. A detector that combines strong accuracy with linear time sequence processing and explicit modeling of sensor topology is therefore well matched to deployment realities, particularly at the network edge where IoT devices typically live.

The Mamba component deserves particular attention because it represents a broader shift in the machine learning community. Since the original Mamba paper appeared, researchers have rapidly adapted the architecture to forecasting and detection tasks, producing variants such as bidirectional Mamba for time series prediction, spatial temporal Mamba for multivariate anomaly detection, and Mamba based foundation models for general forecasting. By embedding Mamba within a graph structure and enhancing its inputs with wavelets, the Beijing team positions their model at the intersection of several active research fronts. The graph element draws on a lineage that runs from graph convolutional networks through temporal graph convolutional networks for traffic prediction to recent topological analysis methods, while the wavelet element revives a classical signal processing tool whose value modern deep learning has increasingly rediscovered.

The study also acknowledges the practical ecosystems in which such models must be validated. The references underpinning the work include widely used benchmarks such as the MIT BIH arrhythmia database for cardiac signals, spacecraft telemetry datasets with expert labeled anomalies, and secure water treatment testbeds used in adversarial cyberattack research. Datasets of this kind have shaped the field by providing realistic fault patterns against which methods can be calibrated, and the competitive advantage the new model demonstrated against baselines on this landscape suggests it generalizes across heterogeneous data regimes rather than being tuned to a single domain. The authors declare no conflict of interest, and the work received support reflected in the supervisory and funding roles of the senior researchers on the team.

Looking ahead, the research points toward anomaly detectors that are simultaneously faster, more aware of physical structure, and more sensitive to the fine grained dynamics of the systems they watch. As billions more devices come online, the volume of time series data will only grow, and the cost of missing a fault will continue to rise, whether measured in damaged equipment, lost revenue, compromised security, or human health. A model that can read a system through its frequencies, its connections, and its history at linear cost offers a template for the next generation of monitoring infrastructure. The Beijing team’s results suggest that the combination of wavelets, graphs, and selective state space models is more than a fashionable assembly of components; it is a coherent answer to the question of how machines can learn what normal looks like in an increasingly connected world, and how quickly they can notice when that normality shatters.

Subject of Research: A wavelet-enhanced Mamba and graph neural network model for detecting anomalies in multivariate time series data from IoT systems.

Article Title: Wavelet-enhanced Mamba and graph network-based model for time series anomaly detection

Article References: Xu, H., Ren, J., Liu, J., Ma, Z., Huang, W., & Niu, J. (2026). Wavelet-enhanced Mamba and graph network-based model for time series anomaly detection. International Journal of Machine Learning and Cybernetics, 17(10), Article 475. https://doi.org/10.1007/s13042-026-03250-x

Image Credits: AI Generated

DOI: 10.1007/s13042-026-03250-x

Keywords: time series, anomaly detection, Mamba, graph neural network, wavelet transform, Internet of Things, machine learning, state space models, multivariate data, deep learning, IoT sensors, sequence modeling

The study, published in the journal Smart Agricultural Technology, tackles a problem that has long been overlooked in precision agriculture: what happens when a farm cannot simply turn on the tap whenever it wants. In many European collective irrigation systems, water is distributed through rotational schedules, with each user granted a fixed volume over a predefined time window. For an individual grower, this translates into a hard operational constraint. When the irrigation system spans multiple sectors growing different crops in different substrates, deciding how to split a limited water budget among competing demands becomes a genuinely difficult combinatorial puzzle.

Lead author Tommaso Adamo and colleagues, including Lucio Colizzi, Giovanni Dimauro, Emanuela Guerriero and Nunzia Lomonte, formalized this challenge as the Water Volume Allocation problem. At the start of each irrigation period, a supply window opens and the farm’s main pump draws water from the external infrastructure, filling an on-farm reservoir. The manager must then commit to a specific volume for every sector before the next decision window arrives. Crucially, the consequences of that commitment cannot be observed at decision time, which makes purely reactive strategies, such as the widely used hysteresis control that triggers irrigation when soil moisture drops below a threshold, structurally inadequate.

To solve the problem, the researchers turned to a paradigm known as Empirical Decision Model Learning, or EDML, which bridges machine learning and mathematical optimization. The idea is elegant in its simplicity. A machine learning model is trained offline on historical sensor data to predict how soil moisture in each sector will respond to any candidate water volume, given the current observable state of the system. This trained model, called the empirical component, is then embedded directly as a set of mathematical constraints inside a Mixed-Integer Linear Programming model, the prescriptive component, allowing an optimization solver to jointly compute the best allocation across all sectors simultaneously.

The machine learning component was built from data collected over roughly 200 days in a real multi-sector greenhouse in the province of Lecce, in southern Italy, under a Mediterranean climate. The facility hosts five heterogeneous sectors: two growing tomatoes in sandy clay loam soil, one growing tomatoes in a soilless substrate of agriperlite and coconut fiber, one growing zucchini, and one growing blueberry. Ground sensors monitored volumetric water content, electrical conductivity and pH, while actuators logged every irrigation event. The researchers enriched this stream with meteorological variables from the ERA5 reanalysis archive, including temperature, humidity, solar radiation and wind speed, creating a rich dataset spanning the full spectrum of agronomic states from water stress to saturation.

After careful preprocessing, which included spike removal, boundary clipping and linear interpolation of gaps caused by the sensors’ energy-saving transmission policy, the team engineered a key feature representing the total water volume delivered over each irrigation period. Cross-validated model selection identified XGBoost gradient boosting regressors as the best predictors for every sector, achieving coefficients of determination between 0.89 and 0.99 in the mineral soil sectors. A multivariate sensitivity analysis confirmed that the models had internalized genuine physical causality: increasing the planned water volume monotonically increased the predicted moisture gain, a property the researchers enforced through monotonic constraints to guarantee the optimizer would receive physically consistent guidance.

The optimization layer then embeds each trained regressor as explicit linear constraints, translating the decision trees of the gradient boosting ensemble into indicator constraints that a solver such as Gurobi can process. The objective is lexicographic: first, the model maximizes agronomic adequacy by steering each sector’s predicted soil moisture toward a target point within the optimal range, defined by thresholds such as the Permanent Wilting Point, Maximum Allowable Depletion and Field Capacity; second, among all solutions achieving that goal, it minimizes total water use, reflecting both the economic incentive of volumetric water pricing and the regulatory imperative of efficient water management under the European Water Framework Directive.

The results were striking. Tested across 50 heterogeneous scenarios, the EDML framework drove crops into the optimal moisture range in 74.4 percent of evaluated instances, the highest success rate among all strategies tested, outperforming fixed-interval irrigation, hysteresis threshold control and a more sophisticated predictive threshold approach. It reduced dangerous saturation events to just 5 instances and kept crops out of the irreversible Danger state in all but 6 cases. Even more remarkable was the water savings: total consumption fell to 4,514 liters, a 79.1 percent reduction compared with fixed-interval irrigation, 50.2 percent compared with predictive threshold control and 38.9 percent compared with hysteresis control, breaking the traditional trade-off in which water-saving reactive strategies compromised crop health.

The framework also proved computationally formidable. In scalability tests scaling the system from 5 to 50 sectors, resolution times remained below 0.3 seconds even for the largest instances, with the solver reaching a certified optimal solution in 100 percent of runs across every tested hardware configuration from one to eight CPU cores. This means the approach is not a laboratory curiosity but a practical tool that can deliver real-time allocation decisions within the operational windows of commercial greenhouses, and potentially scale far beyond the five-sector facility where it was validated.

The researchers emphasize that the empirical qualifier is central to the paradigm: the relationship between water decisions and soil moisture response depends on soil type, substrate and crop physiology, so the model must be learned afresh in each new farming context using locally collected data. Because the training phase is entirely decoupled from the optimization layer, growers can retrain their models with standard machine learning tools as new field data accumulates, without touching the optimization machinery. Future work will extend the framework to open-field deployment, incorporating short-term weather forecasts through stochastic programming to build a genuinely anticipatory, risk-aware allocation policy. For now, the study offers a compelling glimpse of agriculture’s data-driven future, one in which every liter of water is placed exactly where the mathematics of crop physiology says it matters most.

Subject of Research: Empirical decision model learning for multi-sector greenhouse irrigation under water supply restrictions

Article Title: Empirical decision model learning for multi-sector greenhouse irrigation under water supply restrictions

Article References: Adamo, T., Colizzi, L., Dimauro, G., Guerriero, E., & Lomonte, N. (2026). Empirical decision model learning for multi-sector greenhouse irrigation under water supply restrictions. Smart Agricultural Technology, 15, Article 102558. https://doi.org/10.1016/j.atech.2026.102558

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102558

Keywords: greenhouse irrigation, water scarcity, machine learning, mixed-integer linear programming, soil moisture sensors, XGBoost, precision agriculture, EDML, water volume allocation, IoT sensors, smart farming, water footprint