Thanh-Tai Khuu1, a, Thanh-Nghi Doan2,b*, Trung-Duc Pham1,c, Quoc-Anh Nguyen1,d, Thai-Ngoc Luong1,e and Quoc-Vu Nguyen1,f

Keywords: Edge AI, Rice Leaf Disease, Environmental Degradations, Res-UNet & YOLO26m, Static-image Diagnosis.
Abstract. Resource-constrained Edge AI in smart agriculture faces severe performance degradation under environmental and adversarial noise. Addressing this bottleneck, we propose an enhanced dual-chain architecture integrating a Res-UNet defense mechanism with YOLO26m for on-demand static-image diagnosis of four rice leaf diseases. Following a “divide and conquer” strategy, Res-UNet functions as an independent preprocessing shield to restore corrupted images. Thus, YOLO26m detects targets accurately without requiring weight modifications. We empirically validated this approach by expanding 892 rice leaf images into 72 complex noise scenarios. Furthermore, cross-domain benchmarking against 2,000 MS-COCO 2017 and PASCAL VOC 2012 images confirmed the shield’s generalization capacity and resistance to overfitting. Quantitative evaluations demonstrate substantial recovery. While severe degradations drop the baseline mean Average Precision (mAP) to 69.7%, our proposed pipeline successfully restores it to 79.8%. Hardware constraint analysis on mobile edge devices reveals a practical inference time of 4018 ms. Ultimately, accepting this latency trade-off to preserve high diagnostic accuracy provides a highly viable strategy for real-world agricultural deployments.