The deployment of AI in IoT environments demands energy-efficient computational paradigms for resource-constrained edge devices. We present SAUNet, a bio-inspired Spiking Attention U-Net architecture for energy-efficient semantic segmentation. Unlike conventional ANNs using dense continuous-valued activations, SAUNet integrates Leaky Integrate-and-Fire (LIF) neurons with surrogate gradient learning, enabling event-driven temporally-aware computation. SAUNet was evaluated on the PlantDoc leaf disease segmentation dataset across temporal windows ( \(\boldsymbol{T = 6, 8, 10}\) ) against standard Attention U-Net (AUNet). Results show SAUNet \(_{T=10}\) achieves Dice coefficient 0.8062 and IoU 0.6822, while reducing energy consumption by 88% versus AUNet. Temporal analysis confirms progressive prediction refinement over discrete time steps. These findings establish SAUNet as a scalable energy-aware neuromorphic architecture suitable for IoT-enabled educational assessment, medical imaging, agricultural monitoring, and edge AI systems requiring high accuracy with low energy budgets.