Edge-Optimized Melon Leaf Disease Detection Using TensorRT-Accelerated Lightweight YOLO on NVIDIA Jetson Nano
Melon leaf diseases such as anthracnose and downy mildew pose a significant threat to agricultural productivity, with potential yield losses of up to 30-80%. Previous research has demonstrated YOLO’s ability to detect plant diseases with high accuracy; however, most studies have focused on GPU-based server environments without considering the feasibility of deployment on edge devices. This study presents a comprehensive evaluation of four lightweight YOLO model variants from two generations: YOLO11n, YOLO11s, YOLO26n, and YOLO26s. These models were deployed on a Jetson Nano 2GB using TensorRT optimization with FP32, FP16, and INT8 precision formats. The results show that YOLO26n TensorRT FP16 achieved the best latency of 44.97 ms (22.23 FPS), with a speedup of 2.29x over the ONNX GPU baseline and 1.31x over the TensorRT FP32 baseline, and a negligible accuracy difference (mAP50 delta of 0.0014). It has been determined that the Maxwell GPU on the Jetson Nano 2GB does not provide support for hardware-accelerated INT8. Consequently, the performance of TensorRT INT8 is equivalent to that of TensorRT FP32 with an internal speedup of 1.00x. This finding is of critical importance to developers of Maxwell-based edge systems.
Authors:
Kuncahyo Setyo Nugroho, Jason Sebastian Sulistyawan, Mahmud Isnan, Wahyu Sudarno, Bens Pardamean, Prabowo Wahyu Sudarno, Nur Sitha Afrilia, Teddy Suparyanto.
2026 International Conference on Computer, Control, Informatics and its Applications