Publication

Abstract : Weeds remain a persistent problem in today’s agricultural systems. Beyond simply competing with crops for essential resources like light, water, and nutrients, weeds often act as reservoirs for pests and diseases that can harm crop health. This competition and the additional threat from pests can significantly reduce farm produce yield and quality, posing ongoing challenges for farmers. Recent deep learning based studies have made significant progress in spotting weeds automatically. However, problems remain because of changing field conditions, weeds and crops looking alike, and the high computing power needed for real-time use. The goal of this study is to achieve high classification accuracy while giving top priority to inference time. ResNet-50 and EfficientNetB0 models were compared using two improved binary-class datasets: the Weed Crop Dataset and the Weed Detection Dataset. In the classification task, both of these models achieved perfect performance metrics, with accuracy, recall, and precision scores of 1. However, since all models demonstrated identical overall accuracy, accuracy alone is insufficient for comparative evaluation. Therefore, inference time was also analyzed to assess the models’ efficiency. The most efficient inference times were recorded for EfficientNetB0 (4.864s and 2.539s) and ResNet-50 (3.002s and 2.170s). In contrast, Convolutional Neural Networks (CNNs) exhibited longer prediction durations, with times of (5.62s and 5.128s), indicating comparatively slower performance. Among these models, ResNet-50 emerges as the most balanced option, offering an optimal trade-off between computational efficiency and predictive accuracy. These findings support the development of high performing, resource-efficient weed detection systems, contributing to the advancement of precision agriculture and promoting more sustainable agricultural practices.