betaFeedback

WorksComputer Science

Benchmarking the State of AI for Crop Yield Forecasting: A Global Assessment Across Modeling Paradigms

Abstract

Accurate crop yield forecasting is essential for food security, agricultural policy, and commodity markets. Although statistical, process-based, and machine learning approaches have all been proposed, their relative strengths remain poorly understood due to the lack of large-scale, standardized comparisons. We address this gap through a comprehensive benchmark of five modeling paradigms – statistical baselines, process-based models, conventional feature-engineered machine learning, deep sequence models, and tabular foundation models – using the CY-Bench benchmark spanning 63 country-crop datasets (>12,000 administrative regions) across six continents. Beyond conventional pooled accuracy metrics, we introduce a multi-dimensional evaluation framework that separately quantifies overall, spatial, temporal, and anomaly prediction skill. Across multiple crops and countries, feature-engineered machine learning and tabular foundation models achieve the strongest overall predictive performance. Data-driven models reduce NRMSE in 72% of countries growing maize and 78% of countries growing wheat, with median reductions of 8.8% and 11.9%, respectively. Across all modeling paradigms, spatial prediction is substantially easier than temporal and anomaly prediction. While data-driven models markedly improve temporal and anomaly prediction over statistical baselines, accurately forecasting year-to-year yield variability remains challenging. These findings suggest that future progress in large-scale crop yield forecasting will depend on the improved characterization of the environmental and management factors driving inter-annual yield variability, founded on effective representations for crop yield forecasting systems that support agricultural decision-making, early-warning applications, and food security planning

CCS Concepts

  • Applied computing → Agriculture.

ACM Reference Format:

Michiel Kallenberg, Philip Janz, Christoph Jörges, Vageesh Saxena, Pratishtha Poudel, Mohammed Musthafa Rafi, and Ioannis N. Athanasiadis. 2026. Benchmarking the State of AI for Crop Yield Forecasting: A Global Assessment Across Modeling Paradigms. In Proceedings of the 33rd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’27). ACM, New York, NY, USA, 11 pages. https://doi.org/XXXXXXXX.XXXXXXX

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from [email protected].

KDD ’27, San Jose, CA, USA

© 2026 Copyright held by the owner/author(s). Publication rights licensed to ACM.

ACM ISBN 978-1-4503-XXXX-X/2027/08

https://doi.org/XXXXXXX.XXXXXXX

Introduction

Crop yield forecasts are crucial in agricultural science. For example, when drought threatens maize production in East Africa or heat waves reduce wheat yields in South Asia, the speed and accuracy of yield predictions influence whether humanitarian organizations pre-position food aid, governments initiate emergency imports, and commodity markets accurately price risk [30]. However, despite decades of research, crop yield forecasts remain subject to substantial uncertainty arising from systems complexity, data limitations and modeling approaches, limiting their usefulness for agricultural decision-making, especially under a changing climate [52]. The crop yield forecasting community lacks a systematic understanding of which modeling paradigms perform most reliably across diverse geographic regions, environmental conditions, crop types, and stages of the growing season [20, 51]. Furthermore, a wide range of modeling approaches exists for crop yield forecasting, including statistical models, process-based crop models, machine learning, deep learning, and foundation models, as well as hybrid approaches [5, 12, 14, 49]. While several studies report improvements, few works offer results across several crops and countries [40, 41, 42]. Typically, comparisons are restricted to a single country, crop, or modeling paradigm, limiting conclusions about generalizability across diverse agricultural systems [8, 20, 29, 51].

Moreover, crop yield prediction poses a broader challenge for machine learning, as models must generalize across both spatial and temporal dimensions. Yield variability is largely determined by persistent spatial differences in productivity between administrative regions, whereas year-to-year variability within regions is smaller and primarily influenced by weather anomalies and extreme events [11]. Consequently, while a model may demonstrate strong overall predictive performance by replicating long-term spatial yield patterns, it may still fail to capture the inter-annual variability. For example, a model may correctly predict that one region consistently out-yields another because of more favorable soils or management, yet may still fail to predict severe yield losses during a drought year. While such a model achieves high overall accuracy, it provides limited value for early-warning applications that depend on anticipating inter-annual anomalies, demonstrating the importance of sufficient spatial and temporal evaluation schemes.

To address these needs, this research presents a comprehensive benchmark of modeling paradigms for sub-national crop yield forecasting based on the recently introduced CY-Bench dataset [20]. CY-Bench combines standardized environmental predictors with harmonized sub-national yield observations spanning multiple crops, countries, and agro-ecological conditions, providing a common basis for rigorous cross-paradigm evaluation.

The main contributions of this research are threefold. First, we conduct a comprehensive benchmark that compares statistical baselines, process-based models, feature-engineered machine learning, deep sequence models, and tabular foundation models within a unified experimental framework for sub-national crop yield forecasting at near-global coverage. Second, we introduce a multi-dimensional evaluation framework that separately quantifies overall, spatial, temporal, and anomaly prediction skill, revealing complementary strengths that remain hidden when evaluation relies solely on pooled accuracy metrics. Third, we demonstrate that current forecasting performance is primarily limited by the ability to predict inter-annual variability rather than persistent spatial productivity patterns, suggesting that future progress will depend increasingly on learning effective representations of the underlying environmental and management drivers.

Together, these contributions provide a strong empirical foundation for developing crop yield forecasting systems that support agricultural decision-making, early-warning applications, and food security planning.

Related Work

Crop yield forecasting has traditionally relied on statistical regression techniques and process-based crop models that simulate crop growth using biophysical principles and agronomic knowledge [5, 23, 49]. More recently, machine learning methods, including tree-based ensembles [45], recurrent neural networks [10], convolutional neural networks [34], transformer architectures [15], and foundation models [47], have demonstrated increasingly strong predictive performance. Most published studies, however, evaluate only specific families of models [39, 46] and typically focus on a single country [7, 23, 50], crop [4, 29], or modeling paradigm [27, 31]. Consequently, reported improvements are often difficult to generalize across agricultural systems and environmental conditions.

Systematic model intercomparison has long been recognized as an important tool for understanding model strengths, weaknesses, and uncertainty. The Agricultural Model Intercomparison and Improvement Project (AgMIP) and its Global Gridded Crop Model Intercomparison (GGCMI) [31] established standardized protocols to compare process-based crop models under common climate forcing, yielding important insights into structural uncertainty and climate impacts on agricultural production. However, these efforts are limited to process-based crop models and yet do not address comparisons with modern data-driven approaches.

Recent benchmark datasets have established standardized resources for machine-learning-based crop yield forecasting. While some benchmarks emphasize field-scale prediction using high-resolution satellite imagery and multimodal observations [24, 29], others, such as CY-Bench, focus on regional crop yield forecasting across multiple crops and countries. This makes CY-Bench particularly well suited for controlled comparisons of diverse modeling paradigms under identical data conditions. Existing benchmark studies primarily compare machine-learning models with process-based approaches [23, 25], but their scope is typically restricted to specific model families, a single country, or a single crop. To our knowledge, no previous work has systematically compared statistical baselines, process-based models, conventional machine learning, deep sequence models, and tabular foundation models within a unified experimental framework while simultaneously evaluating overall, spatial, temporal, and anomaly prediction skill.

Dataset

We conduct experiments on the CY-Bench dataset [20], a global benchmark that integrates publicly available yield statistics with environmental predictors relevant to crop growth, including weather variables, soil moisture indicators, remote-sensing vegetation indices, and soil properties, for sub-national crop yield forecasting. The dataset covers maize in 36 countries and wheat in 27 countries across six continents, providing yield observations for more than 12,000 sub-national administrative regions. Predictor variables are spatially intersected with crop type maps, aggregated to regional units, and temporally aligned with crop-growing seasons, enabling consistent evaluation across diverse agro-ecological conditions.

To improve robustness, we extend the CY-Bench yield target with a standardized quality-control step. For each country–crop dataset, region–year observations are compared with a locally fitted quadratic yield trend. Observations are flagged as outliers when they exceed the fitted trend by more than three standard deviations, and are removed from both training and evaluation. Consistent with recent work on harmonizing sub-national crop statistics [22], we only filter out positive outliers, as unusually low yields may correspond to genuine crop failures and are therefore retained.

Model paradigms

Figure 1 summarizes the benchmark workflow. We evaluate models from five modeling paradigms: statistical baselines, process-based models, feature-engineered machine learning, (deep) sequence models, and (tabular) foundation models.

Figure 1. Overview of the benchmark. CY-Bench provides harmonized sub-national crop yield observations and environmental predictors for benchmarking statistical, process-based, tabular, and deep sequence models under a common experimental protocol. Performance is evaluated using four complementary dimensions: overall, spatial, temporal, and anomaly prediction.

Statistical baselines

We include two statistical baselines commonly used in crop yield forecasting:

Average yield represents the historical mean yield of an administrative region computed over the training years. This baseline captures persistent regional productivity differences, but ignores temporal variability arising from weather, management, climate trends, or other environmental factors.

Trend yield extends the regional-mean baseline by allowing for gradual changes in regional productivity over time. For each administrative region and forecast year, the model considers yield histories from the training period and searches over window lengths of 5–10 years. For each candidate window, a Mann–Kendall test is applied to assess whether yields exhibit a significant monotonic trend; the window with the strongest evidence of trend is selected. A linear regression of yield on calendar year is then fitted to this window and extrapolated to the target year, capturing long-term technological change, management intensification, and slow climatic shifts. If no significant monotonic trend is detected, the prediction falls back to the regional mean.

Process-based modeling

Process-based crop models simulate plant growth using biophysical representations of crop development, carbon assimilation, and resource limitations [36]. These models are typically driven by meteorological forcing data and simulate crop growth throughout the entire growing season, producing yield estimates at the end of the season once physiological maturity is reached.

As a process-based baseline, we employ the dynamic global vegetation model LPJmL (Lund–Potsdam–Jena managed Land) [48], an agro-ecosystem model that simulates crop growth, natural vegetation dynamics, carbon and water fluxes, and the terrestrial nitrogen cycle to estimate crop yields for different crop functional types (CFT) [48, 54]. Crop growth is simulated from daily meteorological forcing variables, including temperature, precipitation, radiation, and yearly atmospheric CO2CO_2 concentrations. Photosynthesis, respiration, biomass allocation, and phenological development are simulated on a daily time step, with phenology governed by temperature-driven growing degree day (GDD) accumulation. Net primary production is partitioned among plant organs according to CFT-specific allocation rules.

We use LPJmL version 5.9 on a global 0.5° grid. To ensure a fair comparison with the machine learning models, we feed LPJmL with the gridded meteorological inputs of the CY-Bench dataset, derived from the AgERA5 reanalysis [6], including daily temperature, precipitation, solar radiation, and wind speed. Soil properties, including soil type and pH, are obtained from the Harmonized World Soil Database [33], while crop distributions are prescribed using the WorldCereal crop masks [53] included in the CY-Bench raw data. Although LPJmL distinguishes between irrigated and rainfed production systems, all crop masks are interpreted as rainfed to match the assumptions adopted throughout this benchmark. We retain the default LPJmL management configuration to provide a globally consistent, agronomically realistic, process-based baseline.

Finally, LPJmL outputs harvested vegetation carbon, which we convert to crop yield using the standard harvest-index formulation

YLPJmL=Ch(c×dm)×10−2,Y_{\mathrm{LPJmL}} = \frac{C_h}{(c \times dm)} \times10^{-2},

where c=0.45c = 0.45 is the biomass carbon content and dmdm is the CFT-specific dry matter content (0.85 for maize and 0.87 for wheat) [16]. To facilitate a fair comparison with the other forecasting models, we further calibrate the simulated yields by matching the observed regional mean and variance computed from the training years, thereby correcting for systematic biases while preserving interannual variability.

LPJmL has been extensively evaluated within GGCMI [32], making it a well-established process-based baseline. However, its relatively coarse spatial resolution limits its ability to represent sub-grid heterogeneity and capture local yield variability.

Data-driven models

This research evaluates several data-driven models for crop yield forecasting formulated as a point-yield prediction task. Given a set of predictors, including weather variables, soil moisture indicators, remote-sensing vegetation indices, and soil properties, each model learns to predict yearly crop yield of a specific region within a country as a regression problem. The evaluated models differ primarily in their use of temporal information and domain knowledge. Tabular models operate on engineered agronomic features and comprise conventional machine learning and tabular foundation models, whereas sequence models learn representations directly from raw temporal predictors without predefined feature aggregation.

Feature-Engineered Machine Learning Models

Feature-based models incorporate domain knowledge by aggregating the raw time-series predictors into agronomically meaningful summary statistics. These include monthly averages of minimum temperature (tmintmin), maximum temperature (tmaxtmax), mean temperature (tmeantmean), precipitation (precprec), climatic water balance (prec−ET0prec - ET_0), and surface soil moisture (ssmssm). End-of-month values are computed for cumulative Growing Degree Days (GDD), cumulative precipitation, cumulative fraction of absorbed Photosynthetically Active Radiation (fPAR), and cumulative Normalized Difference Vegetation Index (NDVI). Static soil and location covariates are encoded separately. To reduce redundancy and retain the most informative predictors, we perform feature selection using the minimum Redundancy Maximum Relevance (mRMR) [43] criterion during the screening phase. The resulting feature subsets are then used to train Ridge Regression, Random Forest (RF), LightGBM, and XGBoost models.

Foundation Models

Like the conventional machine learning models, tabular foundation models operate on the same engineered feature representation. Unlike conventional machine learning, however, they leverage representations learned during large-scale pre-training on external tabular datasets rather than learning solely from the CY-Bench training data. This pretraining enables them to learn general representations that transfer across downstream prediction tasks, often requiring substantially less domain-specific training data. To ensure a fair comparison, all foundation models evaluated in this study are restricted to the predictors available in CY-Bench and do not use any additional data sources during downstream training or inference.

We evaluate three tabular foundation models: TabICL [44], TabPFN [17], and TabDPT [26], which adapt their pretrained representations to crop yield prediction via in-context learning [9].

Deep Sequence Models

Unlike feature-engineered models, which rely on manually derived agronomic summaries, sequence models learn representations directly from daily weather and remote-sensing observations aligned with the crop calendar. To separate the contribution of temporal representations in different encoders, we distinguish between generic late-fusion encoders and architecture-native encoders. That said, all our neural models follow a common late-fusion architecture consisting of a static context encoder, a temporal encoder, a fusion module, and a regression head. By keeping the static context encoder, fusion module, and regression head identical across models, differences in predictive performance can be attributed primarily to the temporal encoder.

  1. Generic late-fusion encoders: These models provide a controlled comparison of temporal encoding strategies within a common late-fusion architecture. Each variant first tokenizes the daily multivariate sequence using a learnable 1D convolution (kernel size pp, stride 1) followed by max-pooling (window and stride pp), producing a sequence of tokens with embedding dimension DD. During hyperparameter optimization, we consider p∈1,3,5,7,9p \in1, 3, 5, 7, 9, where p=7p=7 corresponds to weekly tokens. We evaluate three variants that differ only in the temporal encoder and pooling rule: LSTM (stacked LSTM, last-token pooling), CNN (temporal Conv1D layers, attention pooling), and Transformer (standard Transformer encoder, attention pooling, with learned sin/cos day-of-year positional features). The temporal embedding hitmp∈RD\mathbf{h}^{\mathrm{tmp}}_i \in\mathbb{R}^{D} is fused with the static context representation before yield prediction.

  1. Architecture-native encoders: We additionally evaluate several state-of-the-art time-series forecasting architectures by integrating their temporal backbones into our late-fusion framework: PatchTST [35], Informer [59], Autoformer [56], the Time-Series Transformer (TST) [2], and the linear decomposition models DLinear and NLinear [57]. PatchTST, Informer, Autoformer, and TST are implemented using the Hugging Face transformers library [55] and retain their native tokenization, positional encoding, and attention mechanisms. DLinear and NLinear preserve their original channel-wise linear formulations. For all architecture-native encoders, we apply mean pooling over the temporal dimension to obtain a fixed-length representation, which is subsequently projected to dimension DD before fusion with the static context representation. Although these architectures were originally developed for multi-step time-series forecasting, we repurpose them here as temporal representation learners for scalar crop yields.

Experimental setup and evaluation

This research implements a two-phase evaluation protocol comprising a Screening Phase and a Walk-Forward Phase. The Screening Phase applies a fixed temporal split: for each country-crop dataset, the last five available years are allocated to the test set, the two preceding years to the validation set, and all earlier years to the training set. As temporal coverage varies across countries, the specific calendar years also differ accordingly. This phase is used to screen candidate model architectures, optimize model-specific hyperparameters, select features for feature-engineered models, and, for neural models, determine the training duration via validation-based early stopping. All model selection decisions are based exclusively on validation performance.

The final evaluation employs a walk-forward evaluation protocol [18]. Beginning with the initial training period defined during the Screening Phase, the training set is expanded by one year after each evaluation, and the model is retrained before generating a forecast for the subsequent year. This process is repeated over the five-year test period, producing five sequential out-of-sample forecasts for each country-crop dataset. Unlike a single hold-out split, the walk-forward protocol mimics the operational deployment of crop yield forecasting systems, where new historical observations become available each season while future information remains unavailable.

To ensure a fair and computationally tractable comparison, all hyperparameters selected during screening remain fixed throughout the walk-forward evaluation, thereby avoiding repeated tuning.

Finally, to address stochastic variation during optimization, all trainable models are evaluated using five independent random seeds in the Walk-Forward Phase. Performance metrics are computed separately for each seed, and their average is reported as the final performance estimate. Deterministic models, such as the statistical baselines and the process-based crop model, are evaluated once.

Hyperparameter optimization

For each country–crop–model combination, we perform hyperparameter optimization during the Screening Phase using Tree-structured Parzen Estimator (TPE) Bayesian optimization implemented in Optuna [1]. Each optimization consists of 30 trials over three groups of model-specific hyperparameters: (i) input representation, namely the number kk of mRMR-selected predictors for feature-based tabular models; (ii) model capacity and architecture, including the embedding dimension and temporal encoder depth for neural models, tree depth and split criteria for tree-based models, and ensemble size for tabular foundation models; and (iii) optimization and regularization, including learning rate, weight decay, batch size, dropout, and training duration for neural models, L2 regularization for ridge regression, and learning rate, subsampling ratio, and regularization parameters for gradient-boosting models.

For the feature-based models, mRMR uses the fixed F-test correlation difference (FCD) criterion [58], in which feature relevance is quantified by the F-statistic and redundancy by pairwise correlations. Hyperparameter optimization selects only the number of retained predictors, kk. During the Walk-Forward Phase, we re-fit mRMR independently at each forecast origin using only the corresponding training data while keeping the selected value of kk fixed. Neural models reuse the early-stopped training duration selected during screening. The statistical baselines and the process-based crop model do not require hyperparameter optimization.

Country-specific modeling

To account for differences in agro-ecological conditions and management practices, we train and tune each model independently for every country. Hyperparameter optimization is performed separately for each country, allowing architectural characteristics, such as the number of neural network layers, to adapt to the available data. Consequently, our evaluation compares modeling paradigms and model families (e.g., LSTMs, Transformers, and tree-based models) rather than fixed architectural configurations.

Multi-dimensional evaluation

Crop yield variability is largely driven by persistent spatial differences in productivity among administrative regions with distinct environmental and management conditions, whereas year-to-year variability within regions is comparatively smaller and primarily reflects interannual weather fluctuations and extreme events. Consequently, a model may achieve strong overall predictive performance by reproducing stable spatial yield patterns while failing to capture the temporal variations. To obtain a more comprehensive understanding of model behavior, this research evaluates predictive skill from four complementary perspectives: overall performance, spatial prediction, temporal prediction, and anomaly prediction.

Overall prediction

Following the existing crop yield research, we evaluate the overall performance using all region-year observations in the test set. This approach assesses our model’s ability to predict absolute yields across both spatial and temporal dimensions using the coefficient of determination (R2R^2) and the normalized root mean squared error (NRMSE), both of which are widely used to evaluate model performance in this field [3, 38].

Spatial prediction

Spatial prediction evaluates whether a model reproduces the geographic distribution of yield across administrative regions within a given growing season. For each test year tt, we compute the Pearson correlation between observed and predicted yields across regions:

rsp(t)=r({yr,t}r∈R,{y^r,t}r∈R),(1)r_{\mathrm{sp}}(t) = r\left(\{y_{r,t}\}_{r \in R}, \{\hat{y}_{r,t}\}_{r \in R}\right), \tag*{(1)}

and report spatial prediction skill as the median of rsp(t)r_{\mathrm{sp}}(t) across all test years.

This metric quantifies how well a model captures persistent differences in relative productivity between administrative regions, independent of absolute yield magnitude.

Temporal prediction

Temporal prediction evaluates whether a model captures year-to-year yield variability within administrative regions. For each region rr, we compute the Pearson correlation between observed and predicted yields across the available test years as follows:

rtm(r)=r({yr,t}t∈T,{y^r,t}t∈T),(2)r_{\mathrm{tm}}(r) = r\left(\{y_{r,t}\}_{t \in T}, \{\hat{y}_{r,t}\}_{t \in T}\right), \tag*{(2)}

and report overall temporal prediction skill as the median of rtm(r)r_{\mathrm{tm}}(r) across all regions.

FamilyModelNRMSE ↓R2R^2 ↑rr ↑rr ↑rr ↑
OverallSpatialTemporalAnomaly
Maize (36 countries; 8,211 regions)
BaselineTrend25.4 [19.7, 33.7]0.18 [-0.33, 0.54]0.71 [0.58, 0.82]-0.27 [-0.53, 0.07]-0.10 [-0.17, 0.02]
Process-basedLPJmL34.1† [23.5, 50.5]-0.31† [-1.05, -0.05]0.49† [0.37, 0.65]0.22* [0.04, 0.49]0.08* [-0.01, 0.28]
Feature MLRF24.8 [17.9, 35.3]0.25 [-0.24, 0.52]0.72 [0.46, 0.82]0.27* [-0.09, 0.50]0.16* [-0.02, 0.25]
FoundationTabICL24.2* [17.0, 33.8]0.25* [-0.13, 0.52]0.73 [0.56, 0.83]0.33* [0.07, 0.61]0.14* [0.04, 0.33]
SequenceTransformer23.0 [16.3, 34.3]0.25 [-0.11, 0.50]0.70 [0.49, 0.80]0.14* [-0.04, 0.29]0.02* [-0.03, 0.19]
Wheat (27 countries; 3,878 regions)
BaselineTrend18.3 [15.4, 28.4]0.03 [-0.28, 0.55]0.73 [0.56, 0.89]-0.46 [-0.55, -0.23]-0.18 [-0.36, -0.03]
Process-basedLPJmL23.5 [17.3, 33.8]0.05 [-0.61, 0.39]0.71 [0.53, 0.84]0.34* [0.01, 0.78]0.18* [0.05, 0.55]
Feature MLXGBoost16.7* [14.1, 25.4]0.31* [0.11, 0.61]0.84 [0.61, 0.91]0.27* [0.03, 0.53]0.09* [0.01, 0.23]
FoundationTabDPT15.9* [14.2, 25.1]0.46* [0.08, 0.63]0.83* [0.68, 0.92]0.22* [-0.03, 0.66]0.13* [0.03, 0.47]
SequenceTST19.5 [15.6, 27.0]0.25 [-0.13, 0.44]0.77 [0.59, 0.90]-0.05* [-0.18, 0.30]-0.01* [-0.09, 0.16]

Table 1. Performance of representative models for each model family. Values are reported as median [IQR] across countries. * significant improvement over Trend; † significant degradation relative to Trend (country bootstrap, B = 10,000). Bold = best.

This metric quantifies how well a model reproduces inter-annual yield fluctuations within regions.

5.3.4 Anomaly prediction: To isolate deviations from long-term regional productivity, we center both observed and predicted yields using the observed regional mean,

yˉr=1∣T∣∑t∈Tyr,t,yr,t′=yr,t−yˉr,y^r,t′=y^r,t−yˉr.(3)\bar{y}_r = \frac{1}{|T|}\sum_{t \in T} y_{r,t}, \qquad y'_{r,t} = y_{r,t} - \bar{y}_r, \qquad\hat{y}'_{r,t} = \hat{y}_{r,t} - \bar{y}_r . \tag*{(3)}

Anomaly skill is quantified as the Pearson correlation computed over all de-meaned region–year observations as follows:

ran=r({yr,t′}r,t,{y^r,t′}r,t).(4)r_{\mathrm{an}} = r\left(\{y'_{r,t}\}_{r,t}, \{\hat{y}'_{r,t}\}_{r,t}\right). \tag*{(4)}

This metric evaluates whether the model correctly predicts years with above- or below-average yields, independently of persistent regional productivity differences.

Overall, we evaluate the spatial, temporal, and anomaly prediction using Pearson’s correlation coefficient, as these analyses focus on agreement in spatial and temporal patterns rather than absolute yield magnitudes.

Results

This section provides only a summary of the results from all trained models. The complete benchmark—including all evaluated models, crops, prediction horizons, performance metrics, qualitative analyses, and explainability visualizations—is available through the accompanying interactive dashboard at https://wurai.github.io/AgML-CY-Bench-dashboard.

Comparison of modeling paradigms

We first compare representative models from each modeling paradigm. Representatives are selected based on the lowest median NRMSE across countries. Table 1 summarizes their performance across the four evaluation dimensions introduced in § 5.3.

Several observations emerge. First, overall predictive performance differs between maize and wheat. For wheat, tabular approaches consistently achieve the strongest overall performance, outperforming the statistical, process-based, and sequence baselines on both NRMSE and R2R^2. For maize, differences between the strongest data-driven paradigms are comparatively modest: the sequence representative achieves the lowest median NRMSE, while the strongest tabular model achieves the highest overall R2R^2. Overall, feature-engineered machine learning and tabular foundation models perform similarly, particularly for maize, while deep sequence models do not consistently outperform either approach.

Second, all model families achieve substantially higher spatial than temporal skill. Finally, anomaly prediction remains challenging across all modeling paradigms, indicating that forecasting departures from a region’s expected productivity continues to be one of the most difficult aspects of sub-national crop yield prediction.

The Trend baseline illustrates why evaluating multiple aspects of predictive skill matters. Despite achieving competitive overall accuracy and strong spatial correlation, Trend exhibits negative temporal and anomaly correlations for both crops. This shows that models can achieve good pooled prediction error largely by reproducing long-term regional productivity patterns, without correctly identifying favorable and unfavorable growing seasons. The strongest tabular models substantially improve temporal and anomaly skill relative to Trend, although anomaly prediction remains considerably more difficult than spatial prediction.

Finally, the process-based LPJmL model exhibits a distinct performance profile. Although its overall predictive accuracy is generally lower than that of the strongest data-driven approaches, it achieves comparatively strong temporal skill, with the highest median temporal (and anomaly) correlation for wheat. This suggests that process-based models capture aspects of environmentally driven inter-annual variability despite their comparatively lower overall predictive accuracy.

Country-level benefit of data-driven models

We assess the consistency of model performance across countries by comparing, for each country, the best-performing data-driven model (feature-engineered, sequence, or foundation) with the best traditional baseline (Trend, Average, or LPJmL). The relative improvement is defined as:

ΔAI=min⁡m∈MtradNRMSE⁡(m)−min⁡m∈MAINRMSE⁡(m)min⁡m∈MtradNRMSE⁡(m)(5)\Delta_{\mathrm{AI}} = \frac{\min_{m \in M_{\mathrm{trad}}} \operatorname{NRMSE}(m) - \min_{m \in M_{\mathrm{AI}}} \operatorname{NRMSE}(m)}{\min_{m \in M_{\mathrm{trad}}} \operatorname{NRMSE}(m)} \tag*{(5)}

Figure 2 summarizes the country-level benefit of data-driven models. For wheat, data-driven models outperform the strongest traditional baseline in 21 of 27 countries (78% (95% CI: 63–93%)), yielding a median NRMSE reduction of 11.9% (95% CI: 4.6–13.8%). For maize, improvements are slightly smaller and more heterogeneous: data-driven models outperform the traditional baselines in 26 of 36 countries (72% (95% CI: 58–86%)), with a median NRMSE reduction of 8.8% (95% CI: 2.3–12.4%).

Relative reduction in NRMSE for maize and wheat

Figure 2. Relative reduction in NRMSE of the best data-driven paradigm compared with the best traditional approach. Positive values indicate improved predictive performance through data-driven modeling.

While the primary benchmark evaluates end-of-season forecasting to enable comparison across all modeling paradigms, Appendix A additionally reports representative-model performance for mid-season forecasting. Although predictive performance generally improves toward the end of the growing season, the supplementary analysis demonstrates that useful crop yield forecasts are already possible by mid-season.

Discussion and conclusion

This benchmark provides several insights into the current state of AI for sub-national crop yield forecasting that extend beyond differences in predictive accuracy between individual models.

First, data-driven approaches generally achieve the strongest predictive performance. Conventional machine learning models trained on engineered agronomic features remain highly competitive with tabular foundation models, which use the same feature representation but benefit from large-scale pretraining. Their similar performance suggests that the engineered representation already captures much of the predictive information available in current large-scale crop yield datasets, leaving little room for improvement through pretraining. In contrast, sequence models, which learn directly from raw daily observations, do not consistently outperform either approach. This suggests that, under the relatively small data regimes available for many country–crop combinations, engineered agronomic features provide a stronger inductive bias than learning temporal representations directly from raw observations.

Second, the benchmark reveals a clear distinction between predicting spatial and temporal variability. Across all modeling paradigms, spatial prediction skill is substantially higher than temporal prediction skill, indicating that reproducing persistent regional productivity differences is considerably easier than predicting inter-annual variability within individual regions. This suggests that current models successfully capture stable differences associated with long-term environmental and agronomic conditions, whereas transient environmental drivers, including weather extremes, and water stress, remain considerably more difficult to model. Since the impact of extreme events depends strongly on their timing relative to crop phenology, further progress will likely require richer observations of rare extreme events and their interaction with crop development. This is particularly important for agricultural early-warning systems, as anticipating year-to-year anomalies is more valuable than reproducing long-term gradients.

The contrasting behavior of the Trend and LPJmL baselines further illustrates this distinction. Trend achieves competitive pooled accuracy despite strongly negative temporal skill, demonstrating that good overall performance can largely be obtained by reproducing persistent productivity gradients without explaining inter-annual variability. Conversely, LPJmL sacrifices overall predictive accuracy while exhibiting comparatively strong temporal skill, suggesting that process-based models capture environmentally driven year-to-year responses despite limitations arising from coarse spatial resolution, global parameterization, and limited biophysical process representation especially for extreme event responses.

Together, these findings demonstrate that pooled accuracy alone provides an incomplete assessment of crop yield forecasting performance, and spatial, temporal and anomaly detection skill needs to be evaluated separately. When model evaluation focuses only on aggregate error, the complementary skills of individual models to estimate crop responses is largely obscured.

The complementary strengths of the data-driven, statistical, and process-based approaches further suggest that hybrid modeling frameworks may provide a promising avenue to improve regional crop yield prediction by combining physically consistent process representations with data-driven learning. Learning process-based residuals or designing physics-informed neural networks constrained with crop physics offers the potential to leverage mechanistic crop dynamics while accounting for observational constraints and non-linear relationships [13, 14, 19, 21, 28, 37].

Third, anomaly prediction remains challenging across all modeling paradigms, indicating that current forecasting systems only partially capture the processes driving inter-annual yield variability. Beyond model comparison, these findings also provide insight into the nature of regional crop yield variability itself: persistent spatial productivity patterns can largely be recovered from currently available predictors, whereas inter-annual fluctuations remain only partially explained. One plausible explanation is that current large-scale benchmark datasets contain only limited information on management practices, irrigation, cultivar selection, pests, diseases, and local extreme events, all of which may contribute to this unexplained variability. This highlights a limitation of current forecasting systems, which rely primarily on weather and vegetation indicators while only implicitly representing many of the processes influencing crop production. Taken together, future progress in large-scale crop yield forecasting will require improved characterization of the environmental and management processes driving inter-annual variability, including Earth observation to better characterize crop phenology and stress dynamics, improved agricultural management information, and advances in modeling that better capture the complex relationships between these factors.

The supplementary analysis indicates that useful yield forecasts are already possible by mid-season. This suggests that much of the information predictive of final yield is already available by mid-season, while the incremental value of late-season observations may be reduced by senescence and inter-annual phenological variability.

Beyond these empirical findings, this work demonstrates the value of standardized benchmarking for agricultural AI. By extending CY-Bench with an additional quality-control layer, a walk-forward evaluation protocol, and a harmonized implementation of diverse modeling paradigms, we enable a more rigorous comparison across fundamentally different forecasting approaches. In particular, driving both LPJmL and the data-driven models with identical meteorological forcing, soil properties, and crop distributions minimizes differences arising from input data and evaluation procedures, allowing more direct comparison of model behavior. Such standardized benchmark frameworks are essential for transparent evaluation of future crop yield forecasting methods.

Finally, the benchmark highlights several priorities for future research. Strong feature-engineered baselines should remain part of future evaluations, as they leverage domain knowledge and continue to rival considerably more complex sequence and foundation models. Greater emphasis should also be placed on improving temporal and anomaly prediction rather than further optimizing already strong spatial prediction. More broadly, continued progress is likely to depend not only on improved learning algorithms, but also on richer observations of crop management and transient environmental processes that better explain inter-annual yield variability. Unlike process-based crop models, which explicitly simulate phenological development and evaluate environmental stresses relative to crop growth stages, data-driven models mostly rely on calendar time and must infer these relationships from data. Therefore, future developments may consider learning representations that capture time-series relative to crop phenological stages and thus better capture stage-specific responses to weather and water stress, offering a promising direction for future crop forecasting models and hybrid approaches.

Limitations and ethical considerations

Several limitations should be noted. First, although CY-Bench covers a large number of countries and two major crops, it does not include management variables such as fertilization, irrigation, cultivar choice, tillage, or pest pressure, all of which influence yield variability. Such data are not readily available at continental scales and across countries, but are typically collected only in local-scale experimental agricultural studies. Second, the process-based benchmark is represented by a single globally parameterized LPJmL configuration and should therefore not be interpreted as representative of all process-based crop models. Finally, although the benchmark spans a diverse collection of modeling paradigms, new foundation and multimodal models continue to emerge and should be incorporated in future benchmark releases.

Crop yield forecasts can inform agricultural planning, food security assessments, and commodity markets. They should therefore be used as decision-support tools rather than as standalone decision-making systems, and their predictions should be interpreted together with agronomic expertise and local knowledge. We hope that our benchmark contributes to more transparent and reproducible evaluations of crop yield forecasting models while encouraging future work on improved data quality, richer environmental and management predictors, and robust evaluation protocols.

Generative AI usage

Generative AI tools were used to support software development and manuscript preparation. Cursor AI assisted with implementation of the dashboard. AI-generated content was reviewed and validated by the authors, who remain responsible for the study design, methodology, results, and conclusions.

Data and code availability

The experiments are conducted on the publicly available CY-Bench dataset [20]. We provide an interactive dashboard with results across datasets, model families, prediction horizons, and evaluation metrics at https://wur-ai.github.io/AgML-CY-Bench-dashboard. The source code for this work is available at https://github.com/WUR-AI/AgML-CY-Bench-Modeling.

Acknowledgments

We thank AgML, the Machine Learning team of the Agricultural Model Intercomparison and Improvement Project. This work was supported by the EU Horizon projects AI Foundation Models in Agricultural Sciences (101293777) and Ambrosia (101181300), and the Digital Europe Programme through AgrifoodTEF (101100622).

References

Mid-season forecasting

The primary benchmark evaluates end-of-season forecasting, allowing comparison across all modeling paradigms. To assess the practical feasibility of operational forecasting before harvest, we additionally evaluate representative models using only predictor information available halfway through the growing season. Table 2 summarizes the change in predictive performance relative to end-of-season forecasting across the four evaluation dimensions.

FamilyModelOverall Δ\DeltaNRMSEOverall ΔR2\Delta R^2Spatial Δr\Delta rTemporal Δr\Delta rAnomaly Δr\Delta r
Maize (36 countries; 8,211 regions)
BaselineTrend0.0 [0.0, 0.0]0.00 [0.00, 0.00]0.00 [0.00, 0.00]0.00 [0.00, 0.00]0.00 [0.00, 0.00]
Process-basedLPJmLN/AN/AN/AN/AN/A
Feature MLRF0.3 [-0.7, 1.3]0.01 [-0.03, 0.09]0.00 [-0.04, 0.02]0.06 [-0.10, 0.16]0.01 [-0.05, 0.07]
FoundationTabICL0.4∗0.4^{*} [-1.4, 2.4]0.03∗0.03^{*} [-0.07, 0.10]0.03 [0.00, 0.07]0.12∗0.12^{*} [-0.04, 0.26]0.03 [-0.03, 0.15]
SequenceTransformer-0.1 [-1.2, 1.8]0.00 [-0.06, 0.14]0.01 [-0.03, 0.02]0.01 [-0.05, 0.13]0.02 [-0.04, 0.07]
Wheat (27 countries; 3,878 regions)
BaselineTrend0.0 [0.0, 0.0]0.00 [0.00, 0.00]0.00 [0.00, 0.00]0.00 [0.00, 0.00]0.00 [0.00, 0.00]
Process-basedLPJmLN/AN/AN/AN/AN/A
Feature MLXGBoost0.1 [-0.4, 1.6]0.00 [-0.02, 0.09]0.00 [-0.01, 0.02]0.09 [-0.08, 0.16]0.03 [-0.05, 0.12]
FoundationTabDPT1.5 [-0.4, 3.2]0.05 [-0.02, 0.22]0.02∗0.02^{*} [-0.01, 0.05]0.13∗0.13^{*} [-0.07, 0.33]0.04∗0.04^{*} [-0.02, 0.24]
SequenceTST1.1∗1.1^{*} [-0.5, 2.5]0.08∗0.08^{*} [-0.03, 0.26]0.01∗0.01^{*} [0.01, 0.05]0.11∗0.11^{*} [-0.08, 0.23]0.02 [-0.09, 0.13]

Table 2. Difference between mid-season (∼50%\sim50\% of the growing season observed) and end-of-season (∼100%\sim100\%) performance for the representative model of each family. Values denote the median Δ\Delta [IQR] across countries. Positive Δ\Delta indicates improved end-of-season performance.

Across both crops, predictive performance generally improves when observations from the complete growing season are available. Nevertheless, the relatively small performance differences indicate that substantial predictive skill is already available at mid-season, suggesting that useful yield forecasts can be produced several months before harvest.

References

  1. [1]Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. 2019. Optuna: A next-generation hyperparameter optimization framework. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining. 2623–2631.arxiv.org/abs/1907.10902
  2. [2]Alexander Alexandrov, Konstantinos Benidis, Michael Bohlke-Schneider, Valentin Flunkert, Jan Gasthaus, Tim Januschowski, Danielle C Maddix, Syama Rangapuram, David Salinas, Jasper Schulz, et al. 2020. Gluonts: Probabilistic and neural time series modeling in python. Journal of Machine Learning Research 21, 116 (2020), 1–6.arxiv.org/abs/1906.05264
  3. [3]S Archana and P Senthil Kumar. 2023. A Survey on Deep Learning Based Crop Yield Prediction. Nature Environment & Pollution Technology 22, 2 (2023).DOI
  4. [4]Muhammad Ashfaq, Imran Khan, Dilawar Shah, Shujaat Ali, and Muhammad Tahir. 2025. Predicting wheat yield using deep learning and multi-source environmental data. Scientific Reports 15, 1 (2025), 26446.DOI
  5. [5]Bruno Basso and Lin Liu. 2019. Seasonal crop yield forecast: Methods, applications, and accuracies. In Advances in Agronomy: Vol. 154. Elsevier, 201–255. doi:10.1016/bs.agron.2018.11.002DOI
  6. [6]H. Boogaard, J. Schubert, A. De Wit, J. Lazebnik, R. Hutjes, and G. Van der Grijn. 2022. Agrometeorological indicators from 1979 to present derived from reanalysis. Climate Data Store - Copernicus Climate Change Service, https://doi.org/10.24381/cds.6c68c9bb.
  7. [7]Evangeline Corcoran, Daniel P Bebber, Stelian Curceac, Natalia Efremova, Azam Lashkari, Andrew Mead, Richard J Morris, Richard FPywell, John W Redhead, and Sebastian E Ahnert. 2026. A comprehensive UK crop yield dataset incorporating satellite, weather, and soil type information. Scientific Data 13, 1 (2026), 491.
  8. [8]Kyle Frankel Davis, Weston Anderson, Steffen Ehrmann, Rafaela Flach, Carsten Meyer, Jonathan Proctor, Deepak Ray, Liangzhi You, Michael Foley, Hervé Kerdiles, et al. 2025. HarvestStat: a global effort towards open and standardized sub-national agricultural data. Environmental Research Letters 20, 5 (2025), 052001.DOI
  9. [9]Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Baobao Chang, et al. 2024. A survey on in-context learning. In Proceedings of the 2024 conference on empirical methods in natural language processing. 1107–1128.
  10. [10]Joshua Fan, Junwen Bai, Zhiyun Li, Ariel Ortiz-Bobea, and Carla P Gomes. 2022. A GNN-RNN approach for harnessing geospatial and temporal information: application to crop yield prediction. In Proceedings of the AAAI conference on artificial intelligence, Vol. 36. 11873–11881.
  11. [11]Katja Frieler, Bernhard Schauberger, Almut Arneth, Juraj Balkovic, James Chrysanthacopoulos, Delphine Deryng, Joshua Elliott, Christian Folberth, Nikolay Khabarov, Christoph Müller, et al. 2017. Understanding the weather signal in national crop-yield variability. Earth’s Future 5, 6 (2017), 605–616.
  12. [12]Yohanne Larissa Gavasso-Rita, Simon Michael Papalexiou, Yanping Li, Amin Elshorbagy, Zhenhua Li, and Corinne Schuster-Wallace. 2023. Crop models and their use in assessing crop production and food security: A review. Food and Energy Security 13, 1 (2023), e503. doi:10.1002/fes3.503
  13. [13]Jingye Han, Liangsheng Shi, Christos Pliyanidis, Qi Yang, and Ioannis N Athanasiadis. 2023. DeepOp: A knowledge-guided machine learning model for rice growth simulation. In 2nd AAAI Workshop on Artificial Intelligence for Agriculture and Food Systems.
  14. [14]Jingye Han, Liangsheng Shi, Qi Yang, Jin Yu, and Ioannis N Athanasiadis. 2025. Knowledge-guided machine learning with multivariate sparse data for crop growth modelling. Field crops research 328 (2025), 109912. doi:10.1016/j.fcr.2025.109912DOI
  15. [15]Adib Hasan, Mardavij Roozbehani, and Munther A Dahleh. 2026. VITA: Variational Pretraining of Transformers for Climate-Robust Crop Yield Forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 38551–38559.
  16. [16]M. I. Hermelink, J. G. Conijn, and R. Dankers. 2021. Modelling Crop Yields and Water Discharge for Ethiopia with LPJmL. Wageningen Plant Research Report WPR-1207 (2021), 32. doi:10.18174/581423DOI
  17. [17]Noah Hollmann, Samuel Müller, Katharina Eggensperger, and Frank Hutter. 2022. TabPFN: A transformer that solves small tabular classification problems in a second. arXiv preprint arXiv:2207.01848 (2022).
  18. [18]Rob J Hyndman and George Athanasopoulos. 2018. Forecasting: principles and practice. OTexts.
  19. [19]C. Jörges and T. Hank. 2026. Regional Seasonal Crop Yield Forecasts Through Hybrid Crop Modeling and Remote Sensing Data Assimilation. In EGU General Assembly 2026 (2026-05-03/2026-05-08). EGU26–4037. doi:10.5194/egusphere-egu26-4037DOI
  20. [20]Michiel Kallenberg, Dilli Paudel, Stella Ofori-Ampofo, Hilmy Baja, Ron van Bree, Aike Potze, Pratishtha Poudel, Abdelrahman Saleh, Weston Anderson, Malte von Bloh, et al. 2026. CY-Bench: a comprehensive benchmark dataset for sub-national crop yield forecasting. Earth System Science Data 18, 6 (2026), 3997–4018.
  21. [21]Michiel GJ Kallenberg, Bernardo Maestrini, Ron van Bree, Paul Ravensbergen, Christos Pliyanidis, Frits van Evert, and Ioannis N Athanasiadis. 2023. Integrating processed-based models and machine learning for crop yield prediction. arXiv preprint arXiv:2307.13466 (2023).arxiv.org/abs/2307.13466
  22. [22]Donghoon Lee, Weston Anderson, Xuan Chen, Frank Davenport, Shraddhanand Shukla, Ritvik Sahajpal, Michael Budde, James Rowland, Jim Verdin, Liangzhi You, et al. 2025. HarvestStat Africa—harmonized subnational crop statistics for sub-Saharan Africa. Scientific Data 12, 1 (2025), 690.
  23. [23]Guoyong Leng and Jim W Hall. 2020. Predicting spatial and temporal variability in crop yields: an inter-comparison of machine learning, regression and process-based models. Environmental Research Letters 15, 4 (2020), 044027.DOI
  24. [24]Fudong Lin, Kaleb Guillot, Summer Crawford, Yihe Zhang, Xu Yuan, and Nian-Feng Tzeng. 2024. An open and large-scale dataset for multi-modal climate change-aware crop yield predictions.
  25. [25]Qingjing Liu, Meijian Yang, Koushan Mohammadi, Dongjin Song, Jinbo Bi, and Guiling Wang. 2022. Machine Learning Crop Yield Models Based on Meteorological Features and Comparison with a Process-Based Model. Artificial Intelligence for the Earth Systems 1 (09 2022), 1–34. doi:10.1175/AIES-D-22-0002.1DOI
  26. [26]Junwei Ma, Valentin Thomas, Rasa Hosseinzadeh, Alex Labach, Jesse Cresswell, Keyvan Golestan, Guangwei Yu, Anthony L Caterini, and Maks Volkovs. 2026. Tabpfn: Scaling tabular foundation models on real data. Advances in Neural Information Processing Systems 38 (2026), 172692–172722.
  27. [27]Yuchi Ma, Yawen Shen, Anu Swatantran, and David L Lobell. 2026. Harvesting Alphath Earth: Benchmarking the geospatial foundation model for agricultural downstream tasks. International journal of Applied Earth Observation and Geoinformation 149 (2026), 105258.DOI
  28. [28]Bernardo Maestrini, Gordan Mamić, Pepijn A J van Oort, Keiji Jindo, Sanja Brdar, Ioannis N Athanasiadis, and Frits K van Evert. 2022. Mixing process-based and data-driven approaches in yield prediction. European journal of agronomy: the journal of the European Society for Agronomy 139, 126569 (2022), 126569. doi:10.1016/j.eja.2022.126569DOI
  29. [29]Miro Miranda, Deepak Pathak, Patrick Helber, Benjamin Bischke, Hiba Najjar, Francisco Mena, Cristhian Sanchez, Akshay Pai, Diego Arenas, Matias Valdenegro-Toro, et al. 2026. Yieldsat: A multimodal benchmark dataset for high-resolution crop yield prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 22920–22930.arxiv.org/abs/2604.00940
  30. [30]Isidro Juan Mirón, Cristina Linares, and Julio Díaz. 2023. The influence of climate change on food production and food safety. Environmental Research 216 (2023), 114674.DOI
  31. [31]Christoph Müller, Joshua Elliott, David Kelly, Almut Arneth, Juraj Balkovic, Philippe Ciais, Delphine Deryng, Christian Folberth, Steven Hoek, Roberto C Izaurralde, et al. 2019. The global gridded crop model intercomparison phase 1 simulation dataset. Scientific data 6, 1 (2019), 50.DOI
  32. [32]Christoph Müller, Joshua Elliott, James Chryssanthacopoulos, Almut Arneth, Juraj Balkovic, Philippe Ciais, Delphine Deryng, Christian Folberth, Michael Glotter, Steven Hoek, Toshichika Iizumi, Roberto C Izaurralde, Curtis Jones, Nikolay Khabarov, Peter Lawrence, Wenfeng Liu, Stefan Olin, Thomas A. M. Pugh, Deepak K. Ray, Ashwan Reddy, Cynthia Rosenzweig, Alex C. Ruane, Gen Sakurai, Erwin Schmid, Rastislav Skalsky, Carol X. Song, Khui Wang, Allard de Wit, and Hong Yang. 2017. Global gridded crop model evaluation: benchmarking, skills, deficiencies and implications. Geoscientific Model Development 10, 4 (2017), 1403–1422. doi:10.5194/gmd-10-1403-2017DOI
  33. [33]Freddy Nachtergaele, Harij Van Velthuizen, Luc Verelst, Niels Batjes, Koos Dijkshoorn, Vincent van Engelen, Guenther Fischer, Arwyn Jones, Luca Montanarella, Monica Petri, et al. 2010. Harmonized world soil database. In Proceedings of the 19th world congress of soil science, soil solutions for a changing world, Brisbane, Australia, Vol. 2010. International Union of Soil Sciences, 34–37.
  34. [34]Petteri Nevalauri, Nathanael Narra, and Tarmo Lipping. 2019. Crop yield prediction with deep convolutional neural networks. Computers and electronics in agriculture 163 (2019), 104859.
  35. [35]Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2022. A time series is worth 64 words: Long-term forecasting with transformers. arXiv preprint arXiv:2211.14730 (2022).arxiv.org/abs/2211.14730
  36. [36]Rogério de Sá Noia-Júnior, Alex C Ruane, Ioannis N Athanasiadis, Frank Ewert, Matthew Tom Harrison, Jonas Jägermeyr, Pierre Martre, Christoph Müller, Taru Palosuo, Montserrat Salmerón, et al. 2025. Crop models for future food systems. One Earth 8, 10 (2025).
  37. [37]Alexandros Oikonomidis, Cagatay Catal, and Ayalew Kassahun. 2022. Hybrid deep learning-based models for crop yield prediction. Applied artificial intelligence 36, 1 (2022), 2031822.
  38. [38]Nidhi Parashar, Prashant Johri, Arfat Ahmad Khan, Nitin Gaur, and Seifedine Kadry. 2024. An Integrated Analysis of Yield Prediction Models: A Comprehensive Review of Advancements and Challenges. Computers, Materials & Continua 80, 1 (2024), 389.DOI
  39. [39]Yashraj Patil, Harikrishnan Ramachandran, Sridhevi Sundararajan, and P Srideviponnal. 2025. Comparative analysis of machine learning models for crop yield prediction across multiple crop types. SN Computer Science 6, 1 (2025), 64.
  40. [40]Dilli Paudel, Hendrik Boogaard, Allard de Wit, Sander Janssen, Sjoukje Osinga, Christos Pyliandis, and Ioannis N Athanasiadis. 2021. Machine learning for large-scale crop yield forecasting. Agricultural Systems 187 (2021), 103016. doi:10.1016/j.agsy.2020.103016DOI
  41. [41]Dilli Paudel, Hendrik Boogaard, Allard de Wit, Marijn van der Velde, Martin Claverie, Luigi Nisini, Sander Janssen, Sjoukje Osinga, and Ioannis N Athanasiadis. 2022. Machine learning for regional crop yield forecasting in Europe. Field crops research 276, 108377 (2022), 108377. doi:10.1016/j.fcr.2021.108377DOI
  42. [42]Dilli Paudel, Diego Marcos, Allard de Wit, Hendrik Boogaard, and Ioannis N Athanasiadis. 2023. A weakly supervised framework for high-resolution crop yield forecasts. Environmental Research Letters 18, 9 (2023), 094062. doi:10.1088/1748-9326/acf50eDOI
  43. [43]Hanchuan Peng, Fuhui Long, and Chris Ding. 2005. Feature selection based on mutual information criteria of max-dependency, max-relevance, and min-redundancy. IEEE Transactions on pattern analysis and machine intelligence 27, 8 (2005), 1226–1238.pubmed.ncbi.nlm.nih.gov/16119262
  44. [44]Jingang Qu, David Holzmüller, Gaël Varoquaux, and Marine Le Morvan. 2025. TabICL: A Tabular Foundation Model for In-Context Learning on Large Data. arXiv preprint arXiv:2502.05564 (2025).arxiv.org/abs/2502.05564
  45. [45]Md Abid Hasan Rafi, Mst Fatematu Johora, Mohima Binte Rasel, Md Dulal Haque, and Md Shahzamal. 2026. Optimized Tree-Based Ensemble Methods with Real-Time IoT-Enabled Framework Incorporating Climatic Parameters for Rice Yield Prediction. SN Computer Science 7, 4 (2026), 346.DOI
  46. [46]Mohammed Musthafa Rafi, Timeihin Aynalide, Baskar Ganapathysubramanian, Soumik Sarkar, Adarsh Krishnamurthy, Chinmay Hegde, and Aditya Balu. 2026. Benchmarking Tabular Foundation Models for Agricultural Yield Prediction. In The First International Workshop on AI in Agriculture.
  47. [47]Filip Sabo, Michele Meroni, Maria Piles, Martin Claverie, Fanie Ferreira, Elna Van Den Berg, Francesco Collignarelli, and Felix Rembold. 2026. From rows to yields: how foundation models for tabular data simplify crop yield prediction. Scientific Reports (2026).DOI
  48. [48]Sibyll Schaphoff, Werner Von Bloh, Anja Rammig, Kirsten Thonicke, Hester Biemans, Matthias Forkel, Dieter Gerten, Jens Heinke, Jonas Jägermeyr, Jürgen Knauer, et al. 2018. LPJmL4–a dynamic global vegetation model with managed land–Part 1: Model description. Geoscientific Model Development 11, 4 (2018), 1343–1375. 10.5194/gmd-11-1343-2018DOI
  49. [49]Bernhard Schauberger, Jonas Jägermeyr, and Christoph Gornott. 2020. A systematic review of local to regional yield forecasting approaches and frequently used data resources. European Journal of Agronomy 120 (2020), 126153. doi:10.1016/j.eja.2020.126153DOI
  50. [50]Amit Kumar Srivastava, Nima Safaei, Saeed Khaki, Gina Lopez, Wenzhi Zeng, Frank Ewert, Thomas Gaiser, and Jaber Rahimi. 2022. Winter wheat yield prediction using convolutional neural networks from environmental and phenological data. Scientific reports 12, 1 (2022), 3215.DOI
  51. [51]Lily-belle Sweet, Ioannis N Athanasiadis, Ron van Bree, Andres Castellano, Pierre Martre, Dilli Paudel, Alex C Ruane, and Jakob Zscheischler. 2025. Transdisciplinary coordination is essential for advancing agricultural modeling with machine learning. One Earth 8, 4 (2025).DOI
  52. [52]M. van der Velde and L. Nisini. 2019. Performance of the MARS-crop yield forecasting system for the European Union: Assessing accuracy, in-season, and year-to-year improvements from 1993 to 2015. Agricultural Systems 168 (2019), 203–212. doi:10.1016/j.agsy.2018.06.009DOI
  53. [53]Kristof Van Tricht, Jeroen Degerickx, Sven Gilliam, Daniele Zanaga, Marjorie Battude, Alex Rossu, Joost Brombacher, Myroslava Lesiv, Juan Carlos Laso Bayas, Santosh Karanam, et al. 2023. WorldCereal: a dynamic open-source system for global-scale, seasonal, and reproducible crop and irrigation mapping. Earth System Science Data 15, 12 (2023), 5491–5515. doi:10.5194/essd-15-5491-2023DOI
  54. [54]Werner Von Bloh, Sibyll Schaphoff, Christof Müller, Susanne Rolinski, Katharina Waha, and Sönke Zaehle. 2018. Implementing the nitrogen cycle into the dynamic global vegetation, hydrology, and crop growth model LPJmL (version 5.0). Geoscientific Model Development 11, 7 (2018), 2789–2812.
  55. [55]Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tanguy Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. HuggingFace’s Transformers: State-of-the-art Natural Language Processing. arXiv:1910.03771 [cs.CL] https://arxiv.org/abs/1910.03771
  56. [56]Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. 2021. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. Advances in neural information processing systems 34 (2021), 22419–22430.arxiv.org/abs/2106.13008
  57. [57]Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. 2022. Are Transformers Effective for Time Series Forecasting? arXiv:2205.13504 [cs.AI] https://arxiv.org/abs/2205.13504
  58. [58]Zhenyu Zhao, Radhika Anand, and Mallory Wang. 2019. Maximum Relevance and Minimum Redundancy Feature Selection Methods for a Marketing Machine Learning Platform. arXiv:1908.05376 [stat.ML] https://arxiv.org/abs/1908.05376
  59. [59]Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wanxie Zhang. 2021. Informer: Beyond efficient transformer for long sequence time-series forecasting. In Proceedings of the AAAI conference on artificial intelligence, Vol. 35. 11106–11115.

Paper details

vlnc:W_b4t6s3x9
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
  • Computer Science

Cursor

View statement
Looking for related papers…
Contents