Abstract: Accurate and large-scale crop yield estimation is often required for global food security in modern agriculture. However, existing models of crop yield estimation are confined by the low interpretability from the blind stacking of multi-source variables. Furthermore, there is insufficient generalization of single-crop models due to feature differences among crop species. Conventional two-dimensional spatial modeling cannot fully capture the structural constraints of three-dimensional terrain on yield formation. In this study, a yield estimation framework was constructed with both physiological mechanism interpretability and cross-crop transfer for maize and soybean, the major staple crops worldwide. A GeoCropNet model was proposed for cross-crop yield estimation. Time-series gross primary production (GPP) data were collected with photosynthetic physiological significance as the inputs. A dual-stream architecture was constructed to extract spatiotemporal features. A three-dimensional geographically weighted regression (3D-GWR) module was introduced in the spatial stream to characterize the spatial non-stationarity of yield using geographic location and elevation information. Meanwhile, a meta-knowledge distillation transfer network (MKDTN) was designed to align bidirectional knowledge between maize and soybean. Validation tests were conducted in the U.S. Corn Belt and the black soil region of Northeast China. The performance was compared with mainstream machine learning and deep transfer learning. The physiological traits revealed that maize and soybean presented highly consistent unimodal GPP accumulation curves during the whole growth period. Their yield-sensitive windows overlapped extremely, providing 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. The results showed that there was stronger robustness with high spatial heterogeneity in field-scale tasks. In bidirectional cross-crop transfer tasks, the 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 information increment window for yield estimation. Stable yield signals of maize and soybean were captured for 24 and 40 days before harvest, respectively. Ablation experiments confirmed that GPP shared a stronger characterization than vegetation near-infrared reflectance (NIRv) in cross-crop transfer scenarios. The 3D-GWR module effectively improved the spatial performance. The t-distributed stochastic neighbor embedding (t-SNE) visualization verified that the framework mapped features of different crops into a unified latent space, thus alleviating the feature deviation in cross-domain estimation. In addition, positive transfer was achieved in all test years in the black soil region of Northeast China, with significantly better performance than the benchmark model, indicating its applicability in fragmented planting scenarios. Physiological indicators were integrated with physical meaning. Three-dimensional geospatial information significantly improved the interpretability and cross-species generalization of yield estimation models. The GeoCropNet framework effectively broke the species barrier of single-crop customized models. The findings can provide technical support for precise yield monitoring at large scale using multi-crop collaboration. Future research can focus on the spatiotemporal downscaling fusion of multi-source remote sensing data using higher-resolution GPP products, particularly for more refined yield estimation at field scale