Abstract: Accurate and large-scale crop yield estimation is critical for global food security and agricultural macro-control. This study aimed to address two key limitations in existing crop yield estimation models: the lack of interpretability caused by blind stacking of multi-source variables, and the insufficient generalization ability of single-crop models due to feature differences among crop species, along with the failure of traditional two-dimensional spatial modeling to capture the structural constraints of three-dimensional terrain on yield formation. This study constructed a yield estimation framework with both physiological mechanism interpretability and cross-crop transfer capability for maize and soybean, the major staple crops worldwide.This study proposed the GeoCropNet model for cross-crop yield estimation. The model took time-series Gross Primary Production (GPP) data with clear photosynthetic physiological significance as the core input, and built a dual-stream spatiotemporal feature extraction architecture. A three-dimensional Geographically Weighted Regression (3D-GWR) module was introduced in the spatial stream to characterize the spatial non-stationarity of yield by integrating geographic location and elevation information. Meanwhile, a Meta-Knowledge Distillation Transfer Network (MKDTN) was designed to achieve bidirectional knowledge alignment between maize and soybean. Validations were conducted in the U.S. Corn Belt and the black soil region of Northeast China, with performance compared against mainstream machine learning and deep transfer learning models.First, the analysis of physiological characteristics revealed that maize and soybean presented highly consistent unimodal GPP accumulation curves during the whole growth period, and their yield-sensitive windows were extremely overlapped, which provided a solid theoretical basis for cross-crop knowledge transfer. In in-domain yield estimation tasks, the GeoCropNet achieved competitive performance with state-of-the-art deep learning models at both county and field scales, and showed stronger robustness in field-scale tasks with high spatial heterogeneity. In bidirectional cross-crop transfer tasks, the proposed model outperformed mainstream transfer learning models including Partial Domain Adaptation Neural Network (PDANN), YieldNet and Adaptive Adversarial Domain Adaptation Neural Network (ADANN) in all test scenarios. The maximum coefficient of determination (R²) of cross-crop yield estimation reached 0.81, and the minimum root mean square error (RMSE) was as low as 359.86 kg/hm². Meanwhile, the model achieved consistent accuracy improvement in all 16 independent test sets compared with its in-domain baseline. Phenological sensitivity analysis identified that the early reproductive growth stage (silking to milk stage for maize, flowering to podding stage for soybean) was the core information increment window for yield estimation, and the model could capture stable yield signals 24 days and 40 days before harvest for maize and soybean, respectively. Ablation experiments confirmed that GPP had stronger characterization ability than vegetation near-infrared reflectance (NIRv) in cross-crop transfer scenarios, and the introduction of 3D-GWR module effectively improved the model’s spatial characterization performance. The t-distributed Stochastic Neighbor Embedding (t-SNE) visualization verified that the framework could map features of different crops into a unified latent space, thus alleviating the feature deviation problem in cross-domain estimation. In addition, the model achieved positive transfer in all test years in the black soil region of Northeast China, with significantly better performance than the benchmark model, which confirmed its applicability in fragmented planting scenarios.This study demonstrated that the integration of physiological indicators with clear physical meaning and three-dimensional geospatial information could significantly improve the interpretability and cross-species generalization performance of deep learning-based yield estimation models. The proposed GeoCropNet framework effectively broke the species barrier of single-crop customized models, and provided technical support for large-scale multi-crop collaborative precise yield monitoring. Future research will focus on the spatiotemporal downscaling fusion of multi-source remote sensing data and the coupling of higher-resolution GPP products to achieve more refined field-scale yield estimation.