Applying county-level crop yield models to individual pixels introduces scale mismatches and aggregation bias. We examined these effects across the winter wheat regions of Henan and Shandong, China, using MODIS gross primary productivity (GPP), daytime and nighttime land surface temperature (LST) time series, and geographic coordinates for 2001–2025. XGBoost and a hybrid convolutional neural network (Hybrid-CNN) were trained using county-level yield statistics for 2001–2020 and independently evaluated on 728 county-year observations for 2021–2024. Three strategies were compared for yield mapping at 500 m resolution: unadjusted pixel predictions, an area-weighted county-level constraint, and quantile mapping (QM) followed by the same constraint. Evaluation covered county-level predictive performance, aggregation bias, within-county variability, and internal consistency with GPP. Validation R² values were 0.749 for XGBoost and 0.770 for Hybrid-CNN but declined to 0.508 and 0.399, respectively, in independent temporal testing, with both models underestimating yield. Aggregating unadjusted pixel predictions to counties produced biases of −1.744 and −0.360 t ha⁻¹, respectively, indicating that county-level predictive accuracy did not directly characterize performance following pixel-level application. Both constrained strategies reduced county-level aggregation residuals to approximately zero. At the same county-level reference yield, adding QM increased the median within-county coefficient of variation (CV) from 2.32% to 10.31% and the median interquartile range (IQR) from 0.187 to 0.826 t ha⁻¹ for XGBoost, representing increases of approximately 344% and 342%, respectively. For Hybrid-CNN, CV increased from 8.95% to 10.66% and IQR from 0.672 to 0.819 t ha⁻¹, corresponding to increases of approximately 19% and 22%. Changes in correlations with GPP remained within ±0.004 for both models. QM thus primarily altered the magnitude of within-county variability, while independent field evidence is still required to determine whether it improves spatial accuracy. The county-constrained strategy provides a baseline for historical yield mapping and regional risk analysis, with QM followed by the county-level constraint serving as a sensitivity comparison. Absolute pixel-level yields, interannual anomalies, and extreme yield-loss signals require further independent validation.