Abstract: This study aimed to develop a phenology-guided crop-yield estimation method that could incorporate crop-development knowledge into multisource time-series learning and account for the different temporal characteristics of remote-sensing and meteorological information. The method was intended to improve cross-year estimation accuracy, early forecasting capability, and interpretability for maize and soybean, while examining its applicability to other major crops and regions. Remote-sensing, meteorological, and official yield records for maize and soybean from 12 states in the United States Corn Belt during 2013–2023 were integrated. A Bidirectional Hierarchical Cross-Attention Generalized Kolmogorov-Arnold Network (BiHCA-GKAN) was developed. Phenological prior masks weighted monthly inputs, bidirectional hierarchical recurrent units encoded temporal dependence, symmetric cross-attention linked crop-condition and environmental information, and generalized Kolmogorov-Arnold modules with narrow and broad grids modeled rapidly varying and slowly varying signals, respectively. Cross-year performance was evaluated using leave-one-year-out validation. BiHCA-GKAN achieved mean coefficients of determination of 0.821 for maize and 0.827 for soybean, with corresponding root mean square errors of 949.8 and 336.9 kg/hm². The best annual coefficients of determination reached 0.840 and 0.856, respectively. Compared with the closest competing models, the mean coefficient of determination increased by 0.007 for maize and 0.005 for soybean, while the corresponding root mean square errors decreased by approximately 2.0% and 1.3%. Ablation experiments showed that the complete architecture produced the best overall performance for both crops. The phenological prior masks improved temporal weighting, symmetric cross-attention strengthened the linkage between remote-sensing and meteorological sequences, and the generalized Kolmogorov-Arnold modules improved nonlinear fitting after multisource interaction. Shapley additive explanations (SHAP) analysis showed that near-infrared, shortwave-infrared, and vegetation near-infrared reflectance features contributed more strongly to the model’s yield-estimation mapping than the other inputs. Meteorological information provided supplementary mapping information related to moisture supply, atmospheric evaporative demand, and drought conditions. The largest attribution magnitudes were concentrated from July to September and remained generally consistent across 2021–2023, indicating that the model relied more strongly on observations acquired during these months. Estimation accuracy increased after maize silking and soybean pod formation and became stable by the maize dent stage and soybean leaf-dropping stage. Reliable estimates were therefore obtained approximately two months before the main harvest period. Spatial predictions reproduced the major high- and low-yield patterns across the Corn Belt, although larger errors occurred in several northern and western transition areas. Additional tests in China yielded mean coefficients of determination of 0.809 for winter wheat, 0.728 for rice, 0.716 for maize, and 0.797 for soybean. These results indicated that the model remained effective across multiple crops and regions, although its stability varied with crop type and regional conditions. The findings showed that combining phenological weighting, bidirectional multisource interaction, and differentiated nonlinear mapping improved crop-yield estimation while retaining interpretable links between model inputs and estimated outputs. The proposed framework supported early yield monitoring and could be applied to multiple crops and regions. Further improvement required dynamically updated phenological masks, higher-resolution observations, and additional soil, cultivar, and field-management information to reduce uncertainty under anomalous seasons and heterogeneous environmental conditions.