Tropical Africa faces severe climate vulnerability compounded by a significant forecasting gap where traditional Numerical Weather Prediction (NWP) models struggle with rapid convective dynamics and sparse ground-level data. This study introduces Agri-MultiNet, a novel multimodal Artificial Intelligence (AI) framework designed to enhance agricultural resilience through impact-based weather forecasting. The system integrates heterogeneous data streams, including high-frequency satellite imagery, real-time IoT sensor networks, and specific agro-genomic indicators, within a unified latent space. The core architecture uses Cross-Modal transformers to align disparate datasets, alongside a hybrid spatiotemporal engine combining Convolutional Neural Networks (CNNs) for cloud morphology, ConvLSTMs for temporal dynamics, and Graph Neural Networks (GNNs) for spatial agro-climatic relationships. By incorporating physics-informed constraints and Bayesian uncertainty quantification, the model bridges the resolution gap between synoptic-scale weather and farm-level biological risk. Quantitative evaluations demonstrate that Agri-MultiNet achieves a 24.1% reduction in rainfall intensity Root Mean Square Error (RMSE) and a significant improvement in the Critical Success Index (CSI), rising to 0.68 compared to 0.42 in traditional systems. Furthermore, the framework successfully captures 78% of seasonal yield variance and provides a Decision Lead-Time Gain (DLTG) of 48 to 72 hours. These advancements translated into an 18% reduction in simulated field-level crop losses by enabling proactive such as optimized irrigation and drainage management. Moving beyond deterministic hazard reporting, Agri-MultiNet interventions operationalizes Risk Translation Theory (RTT) to convert atmospheric signals into actionable indicators like crop stress probability. This research establishes a new benchmark for climate-smart agricultural intelligence in data-sparse regions, offering a scalable pathway toward continental food security. By synthesizing multimodal deep learning with localized environmental stressors, the framework empowers smallholder farmers to transition from reactive disaster response to proactive, predictive resilience in an increasingly volatile climate.