PHYIELD: Towards multi-scale maize yield estimation with limited ground reference data via integrated crop models and phenology-guided deep learning

Accurately estimating crop yields across scales (e.g., regional administrative and field scales) is critical for global food security, given mounting challenges agriculture faces from climate change, population growth, soil degradation, and water scarcity. Remote sensing has been widely used for crop yield estimation with its spatially extensive coverage and temporally continuous observations of crop conditions. However, robust yield estimation across scales remains challenging, particularly under conditions of limited availability of crop yield and crop management data. While process-based crop models can capture complex interactions of plant-soil-atmosphere continuum, their large-scale application is limited by computational cost, parameterization difficulty, and model input (e.g., management) constraints. Empirical data-driven models are more flexible and computationally efficient, but they heavily depend on ground-measured yield labels that are often scarce over space and time. To address these challenges, we propose PHYIELD, a novel hybrid yield estimation framework that integrates systematically calibrated process-based crop modeling and phenology-guided deep learning for maize yield estimation at both county and field scales under yield data constraint conditions. PHYIELD mainly comprises three components: systematic crop model calibration, diversified crop model simulation, and phenology-guided deep learning. Given the challenges and prevalence of sparse yield observations in practice, PHYIELD is uniquely devised with limited ground yield records for model building. Its model performance is systematically evaluated for temporal, spatial, and scale transfers, upon comparison with standalone crop model and deep learning counterpart benchmarks. Results showed that the systematic crop model calibration of PHYIELD, particularly via the calibration for Leaf Area Index (LAI), followed by yield (biomass multiplied by Harvest Index [HI]), is essential for region-varying parameterizations of the crop model before coupling with deep learning. With diversified simulations from calibrated crop model and phenology-guided deep learning, PHYIELD achieved correlation (R) of 0.53–0.73, bias of −1.47–0.96 t ha⁻ 1 , and root mean square error (RMSE) of 1.07–1.68 t ha⁻ 1 in the study region, outperforming both crop model and deep learning benchmarks. Under significant yield data constraint condition (e.g., only 2023 county-level yield data of Illinois available), PHYIELD demonstrated superior temporal transferability over the preceding 20 year period (2003–2022) (R = 0.70, RMSE = 1.76 t ha⁻ 1 ), spatial transfer to surrounding states (R = 0.85, RMSE = 1.06 t ha⁻ 1 ), and scale transfer from county to field levels (R = 0.53–0.76, RMSE = 1.55–1.58 t ha⁻ 1 ). By leveraging the complementary strengths of process-based and data-driven modeling approaches, coupled with systematic parameterization of crop model and incorporation of phenology knowledge, PHYIELD provides a more robust and generalizable solution for multi-scale maize yield estimation, particularly under limited data conditions. It can largely support yield-oriented decision-making in agricultural management, policy development, and food security planning.