Predicting neural activity in higher visual areas (HVAs) in mice during spatial navigation is challenging due to the complex interaction between spatial and temporal processing. One recent two-stream hypothesis for rodents proposed by Saleem (2020) suggests that form-based goals and detailed landmark signals are processed by cortical areas biased towards the central visual field, while self-motion signals are processed in the periphery. Inspired by this, we developed a deep learning model that splits visual input into two streams: a convolutional neural network (CNN) processed the central visual field, and a spatiotemporal feature bank processed the periphery. We trained the model on two-photon calcium imaging data from mice performing a visually guided navigation task in virtual reality. The model outperformed a previous state-of-the-art baseline model in predicting neural responses across V1 and HVAs (p < .05), with correlation scores of up to 0.713 for population coding and 0.307 for single neurons. Saliency map analysis revealed widespread optic flow-modulated neurons. Ablation studies showed that switching two streams (i.e., using fine details from the peripheral field and motion-related spatiotemporal features from the central field) decreased performance. In addition, predicting areas biased toward the central field was significantly more accurate when using the central stream alone compared with using the peripheral stream alone (p < .001). These results suggest that optic flow processing is essential for the periphery, while central vision may specialize in detailed stimuli. Our model aligns with neurophysiological data and provides a new framework for understanding functional specialization in mouse vision during navigation. Future work will focus on a deeper analysis of navigation-relevant variables to understand their contribution to neural activity while extending the model to other tasks and sensory modalities to further capture neural dynamics.