The stakes could hardly be higher. Wheat is a cornerstone of China’s national food security, yet escalating global warming has made extreme climate events considerably more frequent and severe, threatening the stability of production. Traditional tools for forecasting yields have struggled in exactly those years when forecasts matter most. Mechanistic crop models, which simulate plant growth using genotype parameters, weather and soil data, often perform poorly under extreme weather because of heavy data demands and structural rigidity. Statistical models, meanwhile, typically assume linear relationships between vegetation indices and yield, blinding them to the nonlinear dynamics that govern crop responses to stress.

Machine learning approaches such as Random Forest, Support Vector Regression and neural networks have improved on these limits, but they frequently fail to extract spatiotemporal dynamics from complex datasets, and most existing yield studies focus on late growth stages, delaying predictions until it is nearly too late to act. Deep learning architectures have begun to close this gap. Convolutional neural networks excel at extracting spatial features, while Long Short-Term Memory networks, with their memory cells and gating mechanisms, are adept at modeling the time-series character of crop growth. Earlier hybrid CNN-LSTM models outperformed either architecture alone, but they relied mainly on one-dimensional convolutions, ignored static variables such as soil, and—critically—did not account for extreme climate events at all, introducing systematic bias into their predictions.

MBF-HybridNet addresses all three shortcomings at once. The model adopts a multi-branch parallel architecture with three distinct modules: a Dynamic Variables Module that processes daily remote sensing and meteorological data, a Dynamic ECIs Module that handles yearly-scale extreme climate index data, and a Static Variables Module that ingests soil properties. The dynamic module stacks three two-dimensional convolutional layers with asymmetric 1 by 2 kernels, each followed by batch normalization and ReLU activation, and inserts a Self-Attention mechanism after the first two layers to focus on the most informative features. The extracted features are then reshaped into sequences and fed into a two-layer LSTM network with 256 units per layer and dropout to curb overfitting. The static module processes soil organic carbon, cation exchange capacity, pH, sand and clay content through its own convolutional stream. A staged fusion strategy then concatenates these streams through successive fully connected layers to produce the final yield estimate.

A key innovation lies in how the data are prepared. Rather than relying on county-level averages, the team used the Thiessen Polygon method to partition the winter wheat planting area into 288 uniform analysis cells centered on evenly distributed sampling points, preserving fine-scale environmental information while avoiding contamination from non-planting surfaces such as mountains and water bodies. The data were then organized as pseudo-2D image tensors, with rows representing time steps, columns representing feature variables and a third dimension representing the aggregated subregions. The growing season was divided into three progressively cumulative windows: the vegetative growth phase from sowing to pre-jointing, the vegetative-reproductive phase spanning jointing to heading, and the reproductive phase from heading to maturity.

To capture the fingerprint of extreme weather, the researchers computed nine extreme climate indices covering heat, frost and precipitation extremes—hot days, heat stress intensity, consecutive hot days, frost days, cold stress intensity, consecutive cold days, heavy precipitation days, consecutive wet days and consecutive dry days—calculated for each growth stage. Because feeding all 21 growth-stage indices into the model risked dimensionality and overfitting, a genetic algorithm was deployed to select the most informative subset for each stage: three indices for the vegetative phase, five for the vegetative-reproductive window, and eleven for the full season.

The performance gains were substantial. Validated with leave-one-year-out cross-validation across 2004 to 2019, MBF-HybridNet achieved R-squared values of 0.756 to 0.765 across the three cumulative growth stages, with mean absolute percentage errors around 4.2 percent, compared to the baseline LSTM’s R-squared values of 0.652 to 0.671 and errors approaching 4.9 percent. Relative to the baseline, the new model cut root mean square error by up to 55.67 kilograms per hectare and mean absolute error by up to 48.12 kilograms per hectare. An ablation study confirmed that the gains arise from the complementary contributions of spatial representation, self-attention and temporal modeling rather than any single component: CNN alone reached an R-squared of 0.646, adding self-attention lifted it to 0.674, and the full CNN-SA-LSTM stack reached 0.763.

The extreme climate indices proved especially valuable in anomalous years. When the study years were split into normal and extreme groups—with 2006, 2013, 2014 and 2019 flagged as extreme—the model without the indices showed visibly degraded accuracy in extreme years, with R-squared values dropping to around 0.72 to 0.74. Adding all indices raised extreme-year R-squared values to as high as 0.799, and the genetically optimized subsets performed even better, reaching 0.803 in the vegetative-reproductive window while reducing computational cost. Across the full record, the GA-based models improved R-squared by 1.6 to 2.1 percentage points over the index-free model. SHAP interpretability analysis revealed a clear phenological pattern: pre-flowering low-temperature events such as frost days and cold stress intensity dominated early stages, while post-flowering heat and water stress—consecutive hot days, heavy precipitation days and consecutive dry days—took over as the key drivers during grain filling. Wind speed, temperature, precipitation and vegetation indices, particularly solar-induced chlorophyll fluorescence, rounded out the most influential predictors.

Perhaps the most striking result is the model’s early-warning capability. Prediction accuracy improved as seasonal information accumulated but plateaued at the vegetative-reproductive stage, meaning that winter wheat yields can be reasonably estimated approximately 30 days before harvest. The model also corrected a persistent weakness of simpler networks: the tendency to overestimate low yields and underestimate high ones, a bias rooted in the imbalanced distribution of yield samples concentrated in the 5000 to 7000 kilograms per hectare range. Residual analysis showed most county-level errors stayed within plus or minus 400 kilograms per hectare, with the strongest performance in medium- and high-yield areas.

The researchers caution that the framework, built and tested in Qingdao’s temperate monsoon climate, would need regional recalibration elsewhere: extreme-climate thresholds should be adjusted for arid or subtropical zones, growth-stage windows redefined by local phenology, and topographic factors such as elevation and slope added for hillier terrain. Management variables—irrigation, fertilization and cultivar choice—were not explicitly modeled and may explain some residual uncertainty. Still, because every input is drawn from public datasets, the model’s architecture offers a transferable blueprint. As climate extremes intensify, tools like MBF-HybridNet could give farmers and policymakers the lead time they need to protect harvests before the damage is done.

Subject of Research: A multi-branch fusion deep learning model that integrates remote sensing, meteorological, soil and extreme climate index data to estimate winter wheat yields under extreme climate events in Qingdao, China.

Article Title: Develop a multi-branch fusion deep learning model to estimate the winter wheat yields under extreme climate events

Article References: Jiang, X., Kong, D., Zhang, J., Zhang, S., Ma, Z., Yu, L., Yang, S., Bai, Y., Ali, S., & Ullah, H. (2026). Develop a multi-branch fusion deep learning model to estimate the winter wheat yields under extreme climate events. Artificial Intelligence in Agriculture. https://doi.org/10.1016/j.aiia.2026.09.001

Image Credits: AI Generated

DOI: 10.1016/j.aiia.2026.09.001

Keywords: winter wheat, yield prediction, deep learning, extreme climate events, remote sensing, LSTM, convolutional neural network, self-attention, extreme climate indices, Thiessen polygons, food security, genetic algorithm

The research team, led by Long Zhang of Nanjing Agricultural University together with colleagues from industry partner Three Gorges Group, worked at a 20-megawatt agrivoltaic demonstration park in Lishui District, Nanjing, in eastern China. The park, established in 2016 across roughly 47 hectares, generates about 24 million kilowatt-hours per year. Its fixed-support arrays run north-south with south-facing modules installed 2.5 meters above the ground at a 24-degree tilt, covering 53.3 percent of the soil below with panel projections. The researchers spent a full year measuring solar radiation beneath and between the panels using Onset HOBO sensors logging every ten minutes, comparing the field data against an open-field control station to define the daylighting rate, the ratio of light inside the array to light in the open.

To turn those measurements into a design tool, the team built a one-to-one three-dimensional model of the array in SketchUp and imported it into ECOTECT, an environmental simulation package capable of calculating solar radiation, daylighting and shading across complex geometries. The model simplified the panels into uniform rectangular layers of glass, silicon cells and backsheet with measured optical properties, ignored minor shading from diagonal braces, and drew its meteorological boundary conditions from historical China Meteorological Administration data for Nanjing. After a grid independence analysis settled on a 64-by-48 node resolution for the north-south vertical section, the researchers validated the simulation against twelve months of field data from July 2023 to June 2024. The agreement was striking: a coefficient of determination of 0.982, a root mean square error of 2.28, and an average relative error of just 4.9 percent, with no monthly comparison exceeding 8 percent error.

With a trustworthy light model in hand, the team defined the design problem. Four geometric parameters emerged as the truly independent variables an engineer can adjust: array span, panel tilt angle, installation height, and the transverse gap between panel rows. Azimuth was fixed by convention and site orientation, module width was locked by commercial standardization, and coverage ratio was treated as a derived quantity rather than a free variable. The chosen ranges reflected real-world engineering: spans of 8 to 12 meters to accommodate machinery, tilt angles of 18 to 36 degrees, heights of 2.5 to 4.0 meters to clear crops and equipment, and transverse gaps of 0 to 0.6 meters. Three conflicting objectives defined success: maximizing the maximum canopy daylighting rate, maximizing annual energy generation per hectare, and minimizing the coefficient of variation of daylighting, a statistical measure of how unevenly light is distributed across the crop canopy.

The first analytical pass was a single-factor sensitivity study, which revealed that no single knob moves all three objectives in the same direction. Raising the tilt angle from 18 to 36 degrees steadily improved power output and modestly improved light availability. Widening the span improved the daylighting rate substantially but eroded energy generation, because fewer panels fit per hectare. The transverse gap behaved similarly, boosting light penetration up to about 0.4 meters before power losses became steep. Installation height barely mattered for electricity but dominated light uniformity, since taller panels cast longer, softer shadows that blend more evenly across the canopy. Analysis of variance later confirmed these hierarchies: span mattered most for the daylighting rate, the transverse gap and span dominated energy generation, and installation height was the decisive factor for uniformity, with tilt acting as a moderate regulatory parameter throughout.

Because interactions between parameters proved statistically significant, the team then fitted quadratic response surface models using a Box-Behnken design of 29 simulation runs, an efficient scheme that captures linear, interaction and curvature effects while keeping every design point inside the engineering-safe range. The resulting regression models were exceptionally accurate, with coefficients of determination of 0.9978 for daylighting rate, essentially 1.0 for energy generation, and 0.9987 for the coefficient of variation, all significant at well below the 0.0001 probability level. But response surfaces alone could not guarantee a global optimum in such a nonlinear, multi-peaked landscape, so the researchers embedded their surrogate models inside a hybrid evolutionary search. The non-dominated sorting genetic algorithm II, or NSGA-II, provided the global engine, using fast non-dominated sorting, elitist preservation and crowding-distance ranking to maintain a diverse Pareto front of trade-off solutions. In each generation, the most promising layouts seeded an artificial fish swarm algorithm, whose foraging, swarming and following behaviors performed local refinement around those candidates before the improved designs were folded back into the main population.

The hybrid AFSA-NSGA-II search converged on a Pareto set of layouts balancing all three objectives. To pick a single recommended configuration without subjective judgment, the team applied the entropy weight method, which derives objective weights from the information content of the data itself, assigning weights of 0.35 to daylighting rate, 0.27 to energy generation, and 0.38 to uniformity. A TOPSIS ranking, which scores each candidate by its distance from the ideal and anti-ideal solutions, then selected the winner: a tilt angle of 30.1 degrees, a span of 9.1 meters, an installation height of 3.1 meters, and a transverse gap of 0.24 meters. Under this configuration the predicted maximum daylighting rate reached 80.1 percent with a coefficient of variation of 18.3 percent and an energy yield of 1.35 megawatt-hours per hectare. Follow-up simulation confirmed the predictions with relative errors of just 1.7 percent for light availability and 3.2 percent for uniformity, and zero error for power. Compared with the original 24-degree, 8-meter, gap-free design, the optimized layout raised light availability by 4.6 percent and cut light non-uniformity by 12.7 percent while sacrificing only 0.40 megawatt-hours per hectare of electricity.

The spatiotemporal gains carried through the growing seasons. During the overwintering crop period, the optimized array achieved a daylighting rate of 81.6 percent, up 1.9 percent from the original, while the coefficient of variation dropped from 41.9 to 25.9 percent, a 16.0 percent improvement in evenness. In the summer-sown period the daylighting rate reached 81.2 percent, up 2.2 percent, and the coefficient of variation fell 20.5 percent to 15.1 percent. To test whether these simulated benefits translate into grain and legumes, the team ran field trials with winter wheat and peanuts, the dominant rotation crops of the Yangtze River middle and lower reaches, under three panel densities: full density, high density and semi density. Light and yield tracked panel density closely. Under the semi-density layout between panels, wheat yielded 5.4 tonnes per hectare against 6.4 in the open field, a 15.6 percent loss, while the full-density between-panel treatment fell 25.0 percent to 4.8 tonnes. Peanuts showed the same gradient, dropping 20.0 percent under semi-density and 30.0 percent under full density relative to the open-field control of 4.0 tonnes per hectare.

The authors are careful about scope. Their optimum is calibrated to the subtropical monsoon climate of Nanjing and to two specific crops, and they note that shade-tolerant species such as forages, vegetables and certain high-value crops may respond differently to panel density. They also flag inter-annual climate variability and the need to couple the layout framework with crop growth models and multi-year environmental data. Still, the practical message is clear and, for a field often driven by rules of thumb, quietly radical: agrivoltaic design is a genuine multi-objective optimization problem, and treating it as one, with validated light simulation, surrogate modeling and hybrid evolutionary search, can buy meaningfully better growing conditions for a modest and quantified energy cost. As agrivoltaics scales worldwide, frameworks like this one offer engineers a numerical basis for deciding where every panel should sit.

Subject of Research: Multi-objective optimization of photovoltaic array layouts in agrivoltaic systems to balance crop light availability and solar energy generation

Article Title: Multi-objective optimization of photovoltaic array layouts on farmland via a hybrid algorithm framework for enhanced agricultural production and energy generation

Article References: Zhang, L., Geng, X., Ding, H., Cao, K., Wang, L., Chen, H., Deng, L., Wu, C., Xiao, M., & Bao, E. (2026). Multi-objective optimization of photovoltaic array layouts on farmland via a hybrid algorithm framework for enhanced agricultural production and energy generation. Artificial Intelligence in Agriculture. https://doi.org/10.1016/j.aiia.2026.09.002

Image Credits: AI Generated

DOI: 10.1016/j.aiia.2026.09.002

Keywords: agrivoltaics, photovoltaic arrays, multi-objective optimization, NSGA-II, artificial fish swarm algorithm, response surface methodology, light environment simulation, ECOTECT, crop yield, winter wheat, peanuts, solar energy

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