An Optimized Deep Learning Approach for Early Weed Detection in Chili Crop Habitats

Abstract

Plant identification is a significant activity in agriculture, botany, and environmental protection for the precise classification of plants into species to ensure an efficient control of the crops, measurement of diversity, and ecosystems control. Conventional processes of plant classification based on manual identification by specialists are resource-intensive, inefficient, and error-prone. To address these limitations, this research explores Deep Learning (DL)-based approaches for automatic plant classification using chilli weeds as an example. This research evaluated and compared the performance of three Convolutional Neural Network (CNN) architectures—MobileNetV2, ResNet50, and VGG16—trained on a self-collected image dataset from the College of Horticultural Engineering and Food Technology (DSLD CHEFT), Devihosur, Haveri, comprising four plant species. Among the models, MobileNetV2 achieved the highest classification accuracy of 96.6%, outperforming ResNet50 with a 95.0% accuracy, and VGG16 with an 88.0% accuracy. MobileNetV2's lightweight design offers a distinct advantage, which enables efficient inference and makes the model highly suitable for future deployment on edge devices with limited computational resources. This study highlights the potential of CNN-based systems for practical applications in agriculture, such as automated weed detection and precision farming.