Predicting cultivar performance with high accuracy and interpretability is essential for optimizing plant breeding strategies and mitigating the risks posed by climate variability. Crop performance, such as yield, is influenced by complex interactions between genetic factors and environmental conditions (GxE), making it challenging to develop reliable predictive models. The prediction of crop performance in untested environments is an enduring challenge in plant breeding, and is an active area of research with numerous approaches being investigated, such as genomic prediction (Meuwissen et al., 2001; Jannink et al., 2010; Tolhurst et al., 2022), phenomic prediction (Parmley et al., 2019; Winn et al., 2023; Singh et al., 2021; Van der Laan et al., 2025), and crop modeling (Jagtap and Jones, 2002; Messina et al., 2006; Sakamoto, 2020), as well as the integration of these different methods (Technow et al., 2015; Cooper et al., 2020; González et al., 2024; Crossa et al., 2025). Traditional crop modeling approaches, including process-based and statistical models, have been widely used but often require extensive datasets and struggle to capture the intricate, non-linear relationships between variables (Jones et al., 2003; Boote et al., 1996).

Machine learning (ML), particularly deep learning, offers an alternative for yield prediction but has historically faced limitations in interpretability, which is critical for domain experts making breeding and management decisions (Ashfaq et al., 2025; Mahesh and Soundrapandiyan, 2024). Additionally, training ML models requires extensive datasets, which are often costly and time-consuming to generate.

Soybean is one of the most grown crops in North America, and breeding efforts have been extensive for the crop. As part of the breeding efforts for North American soybeans, the annual Uniform Soybean Tests (UST) have been coordinated through the United States Department of Agriculture (USDA) since 1941 (USDA-ARS, 2025a,b). Starting out as yield trials initially, the UST is a partnership with public breeders in universities and government organizations across the US and Canada. These trials are meant to allow public breeding programs and small private programs to test current and experimental varieties in a greater number of new environments within the intended range of adaptation. This allows public breeding programs to expand the area of testing available with less associated costs, and also serve as a valuable source of historical and current soybean data. Because these data are publicly available, they are an affordable option for research in predicting crop performance (Krause et al., 2023). Such large collaborative networks involving multiple breeders and researchers generating large multi-year, multi-environment datasets has been utilized for comprehensive studies (Shook et al., 2021; Wartha et al., 2025; Chiozza et al., 2025), and these valuable datasets provide tremendous opportunities for ML-based model development for agriculture and breeding applications.

Recent advances in attention-based models, particularly transformers, have demonstrated significant improvements in time-series prediction by efficiently capturing long-range dependencies and highlighting critical input features (Vaswani et al., 2017; Singh et al., 2025). Unlike long short-term memory models (LSTMs) (Hochreiter and Schmidhuber, 1997), which process sequences sequentially, transformers utilize self-attention mechanisms to weigh the importance of different time steps and variables simultaneously. This makes them particularly well-suited for yield prediction, where weather variability across the growing season plays a crucial role in determining crop performance (Tan et al., 2018; Shook et al., 2021). Compared to traditional statistical models and deep learning methods such as LSTMs, transformers offer a more flexible and scalable framework for handling complex, multivariate agricultural datasets. The ability to process entire sequences in parallel allows for efficient learning of patterns across years and locations (Shi et al., 2024). Furthermore, these models improves interpretability by explicitly highlighting key weather variables and their temporal importance, aiding plant breeders and agronomists in understanding the drivers of yield variability. Prior studies have utilized traditional and deep learning techniques to model temporal dependencies in yield prediction, there is a need to understand the role(s) of environmental factors and time periods that most influence the critical traits (Lobell et al., 2013), such as soybean seed yield, oil, and protein.

Interpretability plays a vital role in yield prediction by aligning model outputs with agronomic expertise and supporting decision making in breeding programs. This need becomes even more pressing in multi-sensor or multi-source settings, where interpretability has often been cited as a significant gap in current approaches (Ayanlade et al., 2024; Singh et al., 2025). Prior studies have explored both intrinsic and post-hoc interpretability approaches. Intrinsic methods, such as linear regression and decision trees (Monteiro et al., 2022; Dwivedi et al., 2023; Balcan and Sharma, 2024; Torsoni et al., 2023), directly expose feature contributions through their model structures, while post-hoc methods—such as LIME (Ribeiro et al., 2016), SHAP (Lundberg and Lee, 2017), Integrated Gradients (Sundararajan et al., 2017), and DeepLIFT (Shrikumar et al., 2017) approximate local explanations for complex deep models. In recent years, attention-based architectures have also gained prominence as interpretable modeling strategies for sequential and multimodal data, particularly in crop monitoring and yield prediction tasks where temporal attention highlights critical growth stages or stress periods (Xiong et al., 2024; Aviles Toledo et al., 2024; Shook et al., 2021; Wang et al., 2024). However, both intrinsic and post-hoc methods often lack standardized measures of faithfulness, raising concerns about whether their explanations truly reflect the model's decision process rather than coincidental correlations. To address this limitation, evaluation pipelines based on quantitative faithfulness metrics, such as comprehensiveness and sufficiency (DeYoung et al., 2019), and their regression-oriented extension AOPCR (Ozyegen et al., 2022) have been established to assess how model predictions respond to systematic perturbations of the most relevant features. These frameworks enable rigorous evaluation of interpretability methods across domains, providing deeper insight into the reliability of attention- and attribution-based explanations in multi-source agronomic modeling.

In this study, we propose a Transformer-based deep learning model for the prediction of soybean seed yield, oil and protein using historical data from the Uniform Soybean Trials (UST), encompassing the Northern and Southern trials. Our model integrates multidecadal weather records with pedigree information, leveraging recent advances in sequence modeling to improve both predictive accuracy and model interpretability. In addition, by leveraging the self-attention mechanism of Transformer architectures, we identify the most influential time periods and variables during the growing season, providing valuable insights into how environmental factors impact yield, oil and protein levels, offering a valuable tool to potentially mitigate risks and improve agricultural and breeding decision-making. These insights will inform breeding strategies aimed at developing resilient cultivars, ultimately contributing to global food security (Rose et al., 2016). Previous studies that applied attention-based mechanisms to yield prediction (Xiong et al., 2024; Aviles Toledo et al., 2024; Shook et al., 2021), primarily relied on LSTMs and meteorological data without integrating genetic information and crop management practices including pre-planting information. Our work bridges this gap by incorporating a comprehensive dataset including weather, genotype, and management factors, with Transformer-based modeling ensuring a more holistic approach to soybean yield, oil, and protein prediction.

This study presents a comprehensive multimodal frameworks for soybean trait prediction that integrates genotypic relatedness, maturity group, and weather features across multiple locations and years in North America. Previous studies have largely relied on narrow environmental ranges with limited years of data, limited number of cultivars, or single modal data. In contrast, our dataset spans diverse maturity zones and climatic conditions encompassing the soybean growing region of North America, and includes thousands of distinct soybean varieties, which is representative of genetic diversity in soybean breeding programs. The transformer-based architecture leveraged the state of the art attention mechanism, capturing temporal dependencies, while simultaneously enabling cross-modal feature learning in this extensive dataset spanning decades, locations, growing conditions and genotypes.

Earlier research in yield prediction primarily focused on meteorological drivers and vegetation indices derived from remote sensing, often using recurrent neural networks or regression-based ensemble models (Xiong et al., 2024; Schwalbert et al., 2020; Shook et al., 2021; Barbosa Dos Santos et al., 2022; Torsoni et al., 2023; Aviles Toledo et al., 2024; Eugenio et al., 2020). These approaches serve a useful purpose; however, they are less generalizable and lack explainability. Studies have worked on predictions for oil and protein in addition to yield, although these have often been in the context for whole field predictions of a single cultivar (Hernandez et al., 2023; Pereyra et al., 2026), and not the performance of different cultivars as in plant breeding. While these previous plant breeding related approaches have achieved moderate success for yield estimation, their extension to seed composition traits such as oil and protein has been more limited due to data scarcity, and imbalance across environments. Herein, the cooperatively run Northern and Southern Uniform Soybean Test are an excellent community resource, and critical for ML models requiring large datasets, such as Transformers.

The transformer framework presented here captures long-range dependencies among environmental, genetic, and management variables, offering a unified architecture that can be adapted to multiple traits. The observed improvements in predictive accuracy show the importance of explicitly modeling spatio-temporal interactions in the data. Crop development and yield formation are inherently time dependent, with physiological processes responding dynamically to environmental and management changes (Hodges and French, 1985; Chen and Wiatrak, 2010). The transformer's temporal attention enables the model to modify weights on critical growth periods, where weather and soil interactions most strongly influence final outcomes. While our models did not utilize soil features as they were unavailable, future studies are needed to incorporate soil variables in yield prediction (Carroll et al., 2024).

Computational approaches, i.e., Transformer-based models, highlighted different environmental variables carrying higher attention weight in predicting yield. For example, both Northern and Southern latitudes had solar radiation and temperature as most predictive. While this is expected and follows physiological reasoning, it is important to note that our models show a high level of accuracy for genotype level yield prediction simply with weather variables and pedigree at individual locations.

Attention-based interpretability revealed that the variables temperature and solar radiation were most attended during the soybean growing season, confirming their central role in regulating photosynthetic activity during seed development (Hamed et al., 2021; Khaki and Wang, 2019). Air temperature has previously reported to affect seed yield, oil, and protein, with higher temperatures often shortening the growing season, thereby reducing yield and oil content (Piper and Boote, 1999; James E. Board and Modali, 2005; Thenveettil et al., 2024), and reducing seed quality (Egli et al., 2005). Solar radiation is also known to affect soybean development, including impacting biomass, seed yield, and seed composition (Shibles and Weber, 1966; Board and Kahlon, 2011; Bianculli et al., 2016). Although we do not interpret the attention weights as absolute measures of variable importance, they reflect the model's focus conditioned on interactions among each variable over each temporal sequence. Our results indicate how the attention weights reflect the importance of environmental data, and that this new method is able to identify and focus on specific variables which have been identified as important in crop physiology. Together, these analyses demonstrate that the model effectively learned physiologically meaningful associations without being implicitly programmed with pre-assigned weights. Additionally, the use of self-attention provides a mechanism to guide future modeling efforts.

The proposed framework advances the state of trait prediction modeling by enabling generalizable, interpretable, and multi-objective inference on genotypes across traits and locations. Nevertheless, challenges remain. Data imbalance across maturity zones and limited representation of extreme weather events may affect performance under uncertain growing weather conditions. Additionally, while attention mechanisms provide interpretability, they remain correlational rather than causal. Thus, there is a need for an integration with domain-based process models for deeper inference. There is a need for incorporating genomic data, such as SNP markers or geno expression profiles, that can enhance genotype-specific predictions and improve the model's ability to capture G × E interactions at a finer resolution. Another promising direction involves integrating other high-throughput phenotyping data such as hyperspectral indices, or canopy temperature to enrich temporal modeling with more physiological context. Both of these approaches would provide greater information on individual genotypes, allowing for greater capture of the variability caused by unique varieties which is presently lacking in this study. However, generating such datasets will take a concerted effort as these types of data are costly and time consuming to generate. Initial advancements are already noted, for example, a recent SNP dataset from the Nothern Uniform Soybean Tests (Wartha et al., 2025).

In this study, we developed and evaluated a transformer-based deep learning framework for predicting soybean yield across multi-environment trials in the U.S., integrating temporal weather data with genotypic cluster and contextual information. Our approach achieved high predictive performance, with the transformer model outperforming traditional and recurrent baselines in yield prediction. By systematically varying temporal resolutions, we identified that intermediate sampling intervals provided the best balance between temporal fidelity and model efficiency, an insight useful for designing future data embedding strategies in Transformer-based agronomic modeling. We additionally demonstrated that pre-trained temporal representations from yield prediction could be effectively transferred to downstream tasks such as oil and protein prediction, which shows the generalizability of the learned features and highlights a practical strategy for modeling traits with limited labeled data. Leveraging the self-attention mechanism inherent to transformers, our interpretability analysis shows the potential of attention-based models not only as a predictive tool but as a tool to generate actionable insights for breeding and crop management, similar to existing post-hoc interpretability methods. This can serve as an important step toward building explicitly interpretable deep learning models, although this should be further tested on the transferability to other crop species.