The integration of machine learning, edge computing, and Internet of Things technologies is enabling real-time, energy-efficient, and context-aware monitoring in smart agriculture. This paper presents a scalable AIoT framework that combines dual communication protocols, LoRa and NB-IoT, with lightweight machine learning models deployed on resource-constrained embedded platforms. The system adaptively assigns workloads between edge and cloud based on data criticality and network conditions, ensuring low-latency inference and reliable operation. Hardware-aware benchmarking demonstrates that real-time prediction is feasible on ultra-low-power devices, achieving up to 86% accuracy with Decision Tree and Random Forest models, while maintaining low memory footprint and energy consumption. Field experiments show that edge inference consistently reduces latency (average 347–383 ms) and variability compared to cloud modes, and the hybrid communication design mitigates the impact of network contention and harsh wireless conditions. Security evaluations indicate that A Denial-of-Service (DoS), Replay, and Man-in-the-Middle (MITM) attacks are effectively mitigated with minimal overhead. These results highlight the framework’s practical applicability, energy efficiency, and resilience, providing a hardware-agnostic roadmap for deploying AIoT-enabled smart farming systems in diverse agricultural settings.