The study, led by Hasan Mirzakhaninafchi and colleagues, tackled a deceptively difficult problem in irrigation science. Soil water tension, or SWT, measures the suction force roots must overcome to extract water from the soil, and it is widely regarded as one of the most decision-relevant quantities an irrigator can monitor. Unlike volumetric water content, SWT can be interpreted against crop-, soil-, and irrigation-system-specific thresholds. For peanuts grown on the coarse-textured Coastal Plain soils of southern Georgia, the research team drew on prior work at the same site that evaluated trigger levels of 45 and 70 kilopascals as alert thresholds. The core question was whether machine learning could forecast, up to 168 hours in advance, when root-zone tension would cross those critical lines.

To make the sensor data usable for decision-making, the researchers constructed what they call the SOFT root-zone soil water tension series—a smoothed operational target built from Watermark granular matrix sensors installed at roughly 10, 30, and 50 centimeters depth. At each hourly timestamp, the median of the available depth measurements was taken and then smoothed with a trailing three-hour median, ensuring that every SOFT value was constructed only from observations already in hand. The team was explicit that this aggregation is an operational convenience, not a mechanistic model of root water uptake or a direct measure of plant physiological stress. The 45 and 70 kPa values served as study-specific reference thresholds informed by peanut irrigation literature, not universal stress boundaries.

The dataset spanned four peanut growing seasons, from 2022 through 2025, and drew on multiple data streams: multi-depth tension sensors, an on-site weather station providing rainfall, temperature, humidity, solar radiation and reference evapotranspiration, variable-rate irrigation logs, and planting-date records. Postharvest soil texture characterization in 2025 confirmed the sandy profile of the site—sand content ranged from 66 to 90 percent across sampled depths—but texture was used only for site description, never as a model predictor. After rigorous quality control, 66 of 72 monitored plot-season records contributed accepted forecasting windows, ultimately yielding 162,999 analysis windows after an overlap audit removed 2,672 calibration origins whose target intervals bled into validation data.

The modeling lineup read like a catalog of contemporary time-series deep learning. A long short-term memory encoder-decoder, or LSTM-S2S, processed sequences in physical and standardized units. A Temporal Fusion Transformer, or TFT, combined recurrent processing, variable selection, gated residuals, and attention to produce quantile forecasts. PatchTST segmented the input history into overlapping 24-hour patches processed by a transformer encoder. A TCN–TFT hybrid augmented the transformer backbone with two probability channels derived from separate threshold classifiers. All were trained on 2022–2023 data, tuned on an earlier 2024 validation block, calibrated on a later 2024 partition, and finally judged on the entirely held-out 2025 season—a design that guards against the temporal leakage that can inflate performance in agricultural sensor studies.

The headline result was humbling for the machines. Persistence achieved a mean absolute error of 10.29 kPa and a root mean square error of 17.01 kPa on the held-out season. Among the neural forecasters, TFT posted the lowest mean MAE at 10.76 ± 0.98 kPa, and LSTM-S2S the lowest mean RMSE at 17.68 ± 0.98 kPa, but every neural model’s average error exceeded the deterministic baseline. The explanation lies in the physics of soil drying: root-zone tension is strongly autocorrelated, and over much of a seven-day horizon the most recent reading remains an excellent reference trajectory, especially when no major wetting or drying event intervenes. The authors argue that persistence deserves to be treated as a substantive benchmark in all future soil-water forecasting work, not a token comparison.

Yet the story changed when the evaluation shifted from trajectory error to threshold alerting. Because exceedance events were class-imbalanced—16.21 percent of test windows crossed 45 kPa within the alert interval, and only 6.78 percent crossed 70 kPa—the researchers emphasized average precision, a metric that penalizes false alarms more informatively than raw accuracy. Here PatchTST led the neural field, with mean AP of 0.768 ± 0.011 at 45 kPa and 0.714 ± 0.017 at 70 kPa, edging past persistence’s deterministic values of 0.746 and 0.693. The differences were small and interpreted descriptively, but they demonstrate something important: a model can carry higher aggregate trajectory error while still ranking future threshold exceedance more effectively. No single model dominated across trajectory error, alert ranking, and fixed-threshold F1 scores.

Perhaps the most methodologically interesting contribution is the transition-specific evaluation. Standard exceedance metrics can reward models for flagging conditions that are already at or near the threshold—useful for monitoring, but not the same as warning of an impending crossing while the field is still below the line. Restricting evaluation to forecast origins below threshold made the task far harder, with positive prevalence dropping to 6.84 percent at 45 kPa and 3.89 percent at 70 kPa. PatchTST again led the neural models with transition AP of 0.433 ± 0.035 and 0.458 ± 0.034, close to persistence’s 0.409 and 0.418, while LSTM-S2S, TFT, and the hybrid fell well behind. Median lead times for true-positive transition alerts reached 25 hours for PatchTST at 45 kPa and 33 hours for TFT, offering a meaningful window for growers to inspect sensor trends, weigh the rainfall forecast, and plan an irrigation response.

The study also explored a complementary route: a direct temporal convolutional network classifier trained not to reproduce the full tension trajectory but to estimate the probability that the threshold would be crossed within a 48-hour alert horizon, following a six-hour exclusion gap that approximates real-world latency between sensing, processing, and action. Across three training series, this direct classifier achieved mean AP between 0.847 and 0.862 at 45 kPa—numerically the strongest alert discrimination in the study—though PatchTST and persistence retained the edge at 70 kPa. Probability calibration, fitted with temperature scaling followed by isotonic regression or histogram binning on the overlap-purged 2024 partition, produced low expected calibration error, but the authors caution that a calibration slope of 0.498 at 70 kPa shows aggregate statistics can conceal miscalibration across parts of the probability range.

The broader lesson extends well beyond peanut fields. By deliberately separating point forecasting, uncertainty quantification, threshold-exceedance ranking, transition detection, and calibrated probability estimation, the study offers a structured framework for judging machine-learning tools in sensor-based irrigation—and a warning against conflating them. The conformal prediction intervals, with held-out coverage between 91.7 and 92.9 percent against a nominal 90 percent, show that uncertainty can be quantified honestly even when point accuracy resists improvement. The authors are careful about scope: the analysis is a retrospective, season-held-out evaluation at a single research site, not a validated real-time deployment, and it did not test whether alerts improved yield, water use, or economic return. Alerts, they stress, are sensor-derived risk indicators to be weighed alongside current trends, expected rainfall, crop stage, and irrigation-system capacity—not automatic prescriptions to water. In an era when artificial intelligence is often sold as a replacement for judgment, the most viral idea here may be the oldest one: before trusting a sophisticated model, check how it fares against the assumption that tomorrow looks like today.

Subject of Research: Machine learning forecasting of root-zone soil water tension and threshold-exceedance alerting for smart irrigation scheduling in peanut production

Article Title: Decision-oriented root-zone soil water tension forecasting and calibrated threshold-exceedance alerting for smart irrigation in peanut production

Article References: Mirzakhaninafchi, H., Porter, W., Rains, G., Taunton, H., Thompson, S., Tavandashti, A., Warren, A., Wood, B., Kandamali, D., Vargas, A., Porter, E., & Hadi, A. M. (2026). Decision-oriented root-zone soil water tension forecasting and calibrated threshold-exceedance alerting for smart irrigation in peanut production. Smart Agricultural Technology, 15, Article 102584. https://doi.org/10.1016/j.atech.2026.102584

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102584

Keywords: soil water tension, smart irrigation, peanut, deep learning, PatchTST, persistence baseline, threshold alerting, conformal prediction, probability calibration, precision agriculture, temporal fusion transformer, held-out season evaluation

The research team assembled a benchmarking framework that put an unusually broad range of forecasting architectures through their paces under strictly identical experimental conditions. Rather than comparing models across different datasets, different preprocessing pipelines and different evaluation windows — a common weakness in the machine learning literature — the authors trained and tested every model on the same photovoltaic power time-series data, sourced from a publicly available Mendeley dataset of historical generation records. The contenders spanned three generations of forecasting technology: classical recurrent and convolutional deep learning models, modern transformer-based architectures, and the newest class of so-called foundation models, which are large networks pre-trained on vast collections of time series and then adapted to specific forecasting tasks.

In the classical corner sat three workhorses of sequence modeling. The long short-term memory network, or LSTM, uses gated memory cells to decide what information from past time steps should be carried forward, allowing it to capture slow-moving trends such as the daily solar cycle. The gated recurrent unit, or GRU, is a leaner relative that merges some of those gates for computational efficiency. Alongside them ran a one-dimensional convolutional neural network, or 1D CNN, which slides small filters across the input sequence to detect local patterns — sudden ramps in output, for instance — without the sequential processing that makes recurrent networks slow to train. These models have dominated solar forecasting papers for the better part of a decade, and they provided the baseline against which the newer architectures had to justify themselves.

The transformer family brought a fundamentally different mechanism to the table. Instead of reading a sequence step by step, transformers use multi-head attention, a mathematical operation that lets the model directly compare any two points in time and weigh their relevance to the prediction at hand. The study evaluated the Temporal Fusion Transformer, or TFT, an architecture designed for interpretable multi-horizon forecasting that combines attention with gating mechanisms and variable selection, and the iTransformer, a recent variant that reorganizes how the attention operation is applied across the dimensions of multivariate time-series data. The appeal of attention in solar forecasting is intuitive: a cloudy afternoon three days ago may matter more to tomorrow’s forecast than the smooth output of yesterday morning, and attention can learn to lock onto exactly those informative episodes.

The most forward-looking entrants were the foundation models. Time-MoE, introduced at the International Conference on Learning Representations in 2025, is a billion-scale time-series foundation model built on a mixture-of-experts design, in which different specialized sub-networks are activated for different inputs, allowing enormous model capacity without a proportional explosion in computation. Mamba4Cast takes a different route entirely, using state-space models — a mathematical framework for describing how a hidden internal state evolves over time — to achieve efficient zero-shot forecasting, meaning it can predict on data it has never seen without any task-specific training. These models represent the industry’s bet that forecasting, like language translation before it, will eventually be solved by enormous general-purpose networks rather than bespoke per-site models.

The headline result of the benchmark is a clear, graded improvement across architectural generations. The LSTM, the oldest model in the lineup, achieved a skill score of 0.29 and a root-mean-square error, or RMSE, of 0.158. In forecasting parlance, the skill score measures improvement over a naive persistence forecast — the assumption that the next value equals the current one — so 0.29 means the LSTM beat that naive baseline by a modest margin. The transformer-based and foundation-model families both pushed RMSE down to around 0.120 and 0.131 respectively, confirming that attention mechanisms and large-scale pre-training do translate into measurably better solar predictions. The gap may look small in absolute terms, but in grid operations, where forecasts drive billion-dollar scheduling decisions, even percentage-point improvements compound into significant economic and emissions savings.

The outright winner, however, was neither a pure transformer nor a pure foundation model, but a hybrid. The authors’ combined CNN-foundation-transformer architecture achieved a mean absolute error of 0.071, an RMSE of 0.106 and a normalized RMSE of 12.1, corresponding to a skill score of 0.64 against persistence forecasting — more than double the LSTM’s score. The logic of the combination is technically elegant. The convolutional front end extracts local, short-scale features such as abrupt irradiance ramps; the foundation component contributes temporal representations learned from an enormous diversity of time series, giving the model a robust sense of periodicity and seasonality; and the transformer’s attention layers then integrate these multi-scale signals, deciding which features from which time steps matter most for the forecast horizon. Visual analysis of the predictions confirmed that this architecture tracked day-to-day generation peaks and variability patterns with high fidelity, precisely the behavior that matters when operators must decide how much reserve capacity to hold.

Why should a hybrid beat its individual components? The answer likely lies in the complementary inductive biases each block contributes. Recurrent and convolutional layers impose strong structural assumptions about locality and continuity, which helps when data is limited; attention imposes few assumptions and can model long-range dependencies, but typically needs more data to shine; foundation models arrive pre-armed with generalized temporal knowledge that transfers across domains. Solar power data exhibits all of these regimes at once — smooth diurnal cycles, noisy cloud-driven fluctuations, and seasonal drift — so an architecture that layers multiple forms of temporal reasoning appears to capture the phenomenon more completely than any single mechanism. The finding echoes a broader trend in the field, where hybrid CNN-LSTM and CNN-LSTM-transformer designs have repeatedly outperformed monolithic models in prior studies of solar and load forecasting.

The practical implications reach well beyond academic leaderboards. Short-term PV forecasting is a linchpin of grid stability: system operators use these predictions to schedule conventional generation, manage battery storage, and participate in electricity markets. A forecast with a skill score of 0.64, as the best hybrid achieved, means substantially fewer surprise deficits on cloudy days and less curtailment of solar output on bright ones. Moreover, the study’s unified evaluation protocol is itself a contribution. Because every model faced identical data, preprocessing and metrics, the results offer a rare apples-to-apples comparison in a literature notorious for incomparable claims, giving practitioners a defensible basis for choosing architectures rather than relying on cherry-picked benchmarks.

The study also maps the road ahead. Foundation models like Time-MoE and Mamba4Cast performed strongly even though they were not designed specifically for solar data, suggesting that general-purpose temporal pre-training is a powerful starting point that domain-specific fine-tuning could push further. The authors’ work was supported by Al-Quds University’s Najjad Zeenni Faculty of Engineering, and the data and supplementary materials accompanying the paper provide a foundation for replication. As renewable penetration deepens, the contest between forecasting architectures will only intensify — and this benchmark suggests the future belongs not to any single paradigm, but to architectures that know how to make attention, convolution and pre-trained temporal knowledge work together.

Subject of Research: Machine learning architectures for short-term photovoltaic power time-series forecasting

Article Title: A systematic evaluation of classical, transformer-based, and foundation models for photovoltaic power time-series forecasting

Article References: Salman, D., Alzeer, I., & Isayed, W. (2026). A systematic evaluation of classical, transformer-based, and foundation models for photovoltaic power time-series forecasting. Neural Computing and Applications, 38(19), Article 764. https://doi.org/10.1007/s00521-026-12498-x

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12498-x

Keywords: photovoltaic power forecasting, deep learning, transformer models, foundation models, LSTM, time-series forecasting, hybrid neural networks, Temporal Fusion Transformer, Time-MoE, Mamba4Cast, renewable energy, grid stability

The research team, led by Amit Kumar Srivastava with colleagues spanning institutions across Germany, Europe, and India, set out to close four persistent gaps in crop yield forecasting. Process-based crop models, the backbone of systems like the European MARS Crop Yield Forecasting System, are physically interpretable but notoriously difficult to calibrate across diverse agroecological zones, and recent assessments show they systematically underestimate yield losses during compound extremes such as Germany’s 2018 drought. Statistical alternatives like the ABSOLUT model developed at the Potsdam Institute for Climate Impact Research are scalable but locked into predefined functional forms that cannot capture emergent nonlinear interactions between weather and crop physiology. Meanwhile, most deep learning approaches treat weather and static landscape data as separate streams, and their black-box nature undermines the trust of the agronomists and policymakers who must act on their predictions.

CropFusionNet adapts the Temporal Fusion Transformer architecture to the specific demands of agricultural prediction. The model ingests two fundamentally different kinds of information: daily time-varying covariates such as precipitation, sunshine duration, minimum and maximum temperatures, modelled soil moisture and soil temperature, vapour pressure deficit, climatic water balance, reference evapotranspiration, and satellite-derived vegetation indices including NDVI, EVI, FPAR, and LAI; and static covariates such as the Müncheberg Soil Quality Rating, elevation and slope from the Copernicus DEM, and crop-specific irrigated area fractions. Gated Residual Networks and Variable Selection Networks dynamically weigh which features matter at each moment, producing interpretable importance scores, while LSTM layers initialized from static context capture cumulative physiological effects and multi-head self-attention links distant events, such as a drought at planting and heat stress near harvest. Crucially, the model was designed to work with raw daily meteorology rather than pre-aggregated weekly or monthly averages, because the authors showed that aggregation into 8-day, 16-day, or monthly intervals measurably degrades accuracy by smoothing away the short heatwaves and drought spells that disproportionately determine final yield.

The data foundation is unusually comprehensive. District-level yield observations for 397 German districts came from a harmonized dataset covering 1979 to 2021, extended to 2023 using official agricultural statistics, and covering winter wheat, winter barley, and silage maize, which together occupy roughly 55 percent of Germany’s agricultural area and represent about 9.5 million hectares of arable land. Meteorological and soil variables arrived at one-kilometre resolution from the German Weather Service, satellite products were extracted from MODIS Terra via Google Earth Engine and resampled to daily resolution using Savitzky-Golay smoothing and cubic spline interpolation, and phenological observations from the DWD station network defined biologically meaningful growing-season windows. The authors even applied year-specific masking so that when an early harvest followed an extreme year, the model would not ingest irrelevant post-harvest noise into its representations of crop growth.

The performance results are striking. On combined validation and test years from 2019 to 2023, CropFusionNet achieved an R-squared of 0.60 for winter wheat with a mean absolute percentage error of just 8.06 percent, an R-squared of 0.44 for the more challenging winter barley, and an R-squared of 0.75 for silage maize with a correlation of 0.87. Across all three crops it consistently beat a Vanilla LSTM, a Simple Transformer, and a 1D residual convolutional network, recording the lowest normalized root mean square error in every case. Bootstrap resampling with 1000 iterations confirmed these estimates were statistically robust. When benchmarked against the ABSOLUT and MARS operational systems on national average yield predictions for 2018 to 2021, the deep learning framework frequently prevailed, most dramatically during the compound hot-and-dry catastrophe of 2018, where its relative error for winter wheat was 2.64 percent compared with 5.17 percent for ABSOLUT and 5.76 percent for MARS, and its winter barley error was a near-perfect 0.43 percent.

What separates CropFusionNet from a purely statistical triumph, however, is what the model reveals about why it predicts what it does. The variable selection weights show that mean soil quality is the single most important static driver across all three crops, accounting for 0.30 to 0.38 of the total attribution weight, with elevation second, reflecting altitude-driven microclimates in the Central Uplands and Alpine Foreland. Temporally, minimum temperature dominates the winter cereals from autumn establishment through early spring, consistent with known sensitivities to cold stress and vernalization, while vegetation indices take over during heading, flowering, and grain filling. For silage maize, early-season importance concentrates on mean temperature, climatic water balance, and vapour pressure deficit, shifting to EVI and FPAR during peak summer biomass development. Intriguingly, the model assigns high weight to vapour pressure deficit even before planting, plausibly encoding how pre-season atmospheric dryness depletes soil moisture and conditions germination prospects.

Perhaps most remarkable is what happens inside the model’s latent space when researchers project its internal embeddings onto principal components. The catastrophic drought years of 2003 and 2018, along with the 2022 summer drought for maize, cluster unmistakably at the extreme negative end of the first principal component across all three crops, while bountiful years like 2014 sit at the opposite pole. The response is also asymmetric: for silage maize, negative yield extremes shift the centroid by minus 12.32 along PC1, more than half again as far as positive extremes shift in the other direction, evidence that the model has genuinely encoded the physiological signature of stress rather than simply regressing toward the mean. Recast as a three-class early warning problem, the model correctly identified low, normal, and high yield tiers well above the random baseline of 0.33, with overall accuracies of 0.64, 0.59, and 0.71 for wheat, barley, and maize respectively, and severe low-versus-high misclassifications were exceedingly rare.

The practical implications extend to when forecasts can be trusted. Lead-time analysis shows winter wheat accuracy improves sharply about 60 days before its late-July harvest, stabilizing near an RMSE of 0.74 tonnes per hectare, while winter barley needs roughly 40 to 50 days of runway. Silage maize, a spring crop with a compressed growing window, proved strikingly predictable early: errors fell below 5 tonnes per hectare a full 72 days before the late-September harvest. This divergence matters operationally, because it means maize-based early warnings can be issued reliably by mid-July, whereas winter cereals demand frequent updates through their sensitive late-spring phenological stages. The authors are candid about limitations: prediction intervals proved somewhat too narrow during extreme years, district-level aggregation obscures sub-district heterogeneity, dynamic management practices like fertilization and cultivar choice are not explicitly modelled, and the framework, trained solely on German conditions, will require regional fine-tuning elsewhere.

Even so, the study represents a meaningful shift in how agricultural AI is built and judged. Rather than treating interpretability as a post-hoc add-on, CropFusionNet bakes transparency into its architecture, letting an agronomist trace a predicted yield deficit back to, say, an anomalous vapour pressure deficit spike during flowering. The spatial maps of feature importance could guide soil conservation subsidies, insurance premium design, and drought-resilient cultivar deployment to the regions where landscape constraints amplify climate vulnerability. The code is openly available on GitHub, and the authors frame their contribution as a call for deep learning in agriculture to reflect underlying biophysical system dynamics rather than merely chasing accuracy. In an era when a single compound extreme can destabilize regional food systems within one growing season, a forecasting tool that is simultaneously fast, honest about its uncertainty, and legible to the people who must act on it may prove one of the most consequential applications of artificial intelligence to climate adaptation yet.

Subject of Research: Interpretable deep learning for uncertainty-aware district-level crop yield forecasting in Germany

Article Title: CropFusionNet: an interpretable deep learning framework for uncertainty-aware crop yield forecasting across Germany

Article References: Srivastava, A. K., Halder, K., Lopez, G., Muduchuru, K., Barbosa, L. A. P., Rahaman, K. J., Behrend, D., Han, L., Nendel, C., Zhao, G., Gaiser, T., Singh, M., Lanka, K., Han, J., Athanasiadis, I. N., Maerker, M., Zeng, W., Alsafadi, K., Rahimi, J., & Ewert, F. (2026). CropFusionNet: an interpretable deep learning framework for uncertainty-aware crop yield forecasting across Germany. Artificial Intelligence in Agriculture. https://doi.org/10.1016/j.aiia.2026.08.016

Image Credits: AI Generated

DOI: 10.1016/j.aiia.2026.08.016

Keywords: CropFusionNet, crop yield forecasting, deep learning, Temporal Fusion Transformer, Germany, winter wheat, silage maize, climate extremes, explainable AI, remote sensing, drought, agriculture