Smart Agriculture in the Fight against Weeds: Analyzing the Impact of Image Quality on Deep Learning Performance

Abstract

Weed management is a critical challenge for sustainable agriculture, driving the adoption of Smart Farming solutions that integrate the Internet of Things (IoT) and deep learning. While significant research focuses on improving object detection algorithms, the influence of data quality from IoT perception devices on system performance remains underexplored. This paper presents a holistic study of an IoT-aligned weed detection framework, using the YOLOv8 model to investigate the impact of image data quality versus quantity. We train and evaluate two identical YOLOv8n models on contrasting datasets: a high-quality, smaller dataset (D1, n=512) and a larger, lower-quality dataset (D2, n=5,061). Our results show a decisive advantage for data quality: the model trained on D1 achieved a mean Average Precision (mAP@50) of 0.90, significantly outperforming the model trained on D2 (mAP@50 of 0.82), alongside higher precision (0.88 vs. 0.74). This empirical evidence underscores that for robust IoT-based detection systems, investing in high-fidelity data acquisition at the Perception Layer is more effective than merely amassing larger volumes of data. The findings offer a practical design principle for developing reliable and efficient smart agriculture solutions, emphasizing the need for system-level optimization beyond algorithmic choice.