Ioannis Athanasiadis bio photo

Ioannis Athanasiadis

Professor and Chair of Artificial Intelligence
Wageningen University & Research

University Page LinkedIn Google Scholar ORCID ACM DL DBLP Web of Science Twitter Email

Corn yield estimation under extreme climate stress with knowledge-encoded deep learning

X. Xiong, R. Zhong, H. Jiang, I. Athanasiadis, Y. Yang, L. Zhu, T. Lin
Abstract

Accurately estimating crop yield under climate stress is vital for global food security, particularly as extreme weather events become more frequent. Data-driven models are increasingly adopted for enhancing yield estimation. They benefit from the effective learning of crop response to the environment from vast amounts of remote sensing and meteorological data. Extreme climate stress conditions that have few yield labels available failed to train these models for modeling crop-stress interactions. Crop responses to such extreme climate stress could exhibit significant delays and sensitivities that can be captured through remote sensing observations. However, this knowledge is not sufficiently utilized in data-driven yield estimation models to address the lack of labels under extreme stress. This study employs attention mechanisms to explicitly encode the time lag effect and phenology sensitivity within a deep learning framework for predicting crop yield under extreme climate stress. The framework consists of two-stream structure separately receiving climate and remote sensing data, with each aggregating input time series into multitemporal feature embeddings. A time-lag-encoded cross attention fuses feature embeddings between climate and remote sensing streams, while phenology-sensitivity-guided linear attention is applied atop the framework for processing these time-lag-encoded features. The proposed model is evaluated across nine Midwestern states within the US Corn Belt at the county level from 2006 to 2023, simulating extreme climate stress scenarios with limited samples. The time lag analysis indicates an average lag of approximately 45 days between the extreme stress event and the maximum vegetation decay event, revealing that such extreme events lead to delayed reductions in crop greenness. General model evaluation results demonstrate that the knowledge-encoded two-stream model (RMSE = 1.17 Mg ha−1) outperforms both the feature-stacking-based two-stream model (RMSE = 1.43 Mg ha−1) and the random forest model (RMSE = 1.68 Mg ha−1) under conditions of extreme climate stress. Model ablation results show that cross attention and time-lag knowledge significantly improve model accuracy compared to direct sum of features, suggesting knowledge-encoded data fusion is more effective than simply summing multi-source input data. The incorporation of time lag-encoded cross-attention mechanisms also facilitates the identification of distinct time lag patterns since extreme climate stress happened, thereby enhancing the model’s interpretability and providing insights into the interactions between environmental factors and crop responses. In-season analysis reveals that the time lag-encoded model captures extreme stress events once they occur, enabling accurate yield predictions up to 8 weeks in advance. The spatial–temporal transferability experiment shows that the knowledge-encoded two-stream model outperforms baseline models across counties from 2013 to 2023. Notably, the proposed model achieves substantial accuracy gains in regions experiencing extreme heat stress, and it also maintains robust performance across most years. Overall, the time lag knowledge could be extended to other forms of environmental stress as long as it’s captured by multi-source data. The cross-attention mechanism as a basic unit enables integration with more knowledge to improve the modeling of complex biomass accumulation and yield formation.

Download full text in pdf format

cover image Published as:
X. Xiong, R. Zhong, H. Jiang, I. Athanasiadis, Y. Yang, L. Zhu, T. Lin, Corn yield estimation under extreme climate stress with knowledge-encoded deep learning, ISPRS Journal of Photogrammetry and Remote Sensing, 231:101-118, 2026, Elsevier, doi:10.1016/j.isprsjprs.2025.10.020.

You might also enjoy (View all publications)