Real-Time Okra Crop Detection among Weeds Using SSD-MobileNet on an Embedded Platform

This study presents a deep learning-based machine vision system for real-time detection of okra crops among weeds using the SSD-MobileNet-v1 architecture deployed on the NVIDIA Jetson Orin Nano platform. The primary objective was to develop a computationally efficient and lightweight detection framework suitable for embedded agricultural applications. A dataset comprising 500 images was acquired under natural field conditions and annotated using LabelImg (version 1.8.1) with bounding boxes to identify okra plants and weed instances. The model was trained for 250 epochs with a batch size of one to ensure stable convergence under limited computational resources. The SSD-MobileNet-v1 network enables multi-scale feature extraction and predicts bounding boxes with confidence scores for accurate real-time detection. The performance of the system was evaluated under different combinations of crop spacing (150, 300 and 450 mm), operational speed (1.5, 2.0 and 2.5 km h⁻¹) and camera height (150, 300 and 450 mm) using a split-plot experimental design with three replications. The results revealed that all factors significantly influenced detection performance, with strong interaction effects indicating non-linear system behaviour. The optimal performance was achieved at 300 mm crop spacing, 1.5 km h⁻¹ operational speed and 300 mm camera height, where the system attained a maximum accuracy of approximately 92%, along with high precision (up to 0.97), recall (up to 1.00) and F1 score (up to 0.96). In contrast, higher operational speed (2.5 km h⁻¹) and camera height (450 mm) resulted in increased false positives and false negatives, reducing accuracy to 57–60%. The developed models exhibited high predictive capability (R² > 0.95) with low variability, confirming robustness under varying field conditions. Overall, the study demonstrates that integrating deep learning with embedded systems provides an efficient and reliable solution for real-time crop detection and site-specific weed management, with significant potential for precision agriculture applications.