Edge Artificial Intelligence for Sustainable Smart Farming:Architecture, Lightweight Segmentation, and Field Validation

This paper presents a comprehensive framework for deploying artificial intelligence at the edge for precision agriculture, addressing critical challenges in water efficiency, crop health monitoring, and real-time decision-making in resource-constrained farming environments. We consolidate three integrated research contributions: (1) a terminal–edge–cloud reference architecture for smart farming integrating sensor fusion, vision analytics, and adaptive irrigation control; (2) lightweight image segmentation models combining MobileNetV3 encoders with modified U-Net decoders, achieving 85% mIoU while maintaining 12–15 FPS on Raspberry Pi and Jetson Nano devices; and (3) cloud–edge collaborative semantic segmentation for crop disease detection on four horticultural crops (mango, sweet orange, chilli, tomato) with Raspberry Pi deployment and performance diagnostics.

The proposed systems achieve water savings of 15–30% through edge-based irrigation control, reduce pesticide applications by 31% through intelligent disease detection, improve crop yield by 10.3%, and deploy at <USD 85 per node—6–60× cheaper than commercial solutions. Edge AI enables continuous operation with 97% uptime despite intermittent connectivity, powered by solar energy at 34 mW. Key innovations include hybrid loss functions balancing pixel-level accuracy with computational efficiency, knowledge distillation for model compression, INT8 quantization for resource-constrained inference, and federated learning for privacy-preserving multi-farm collaboration.The work addresses critical gaps in edge deployment, real-farm generalization, and practical cost-benefit analysis, bridging the divide between advanced AI algorithms and sustainable, accessible agricultural practice.