Analyzing the performance of deep convolutional neural network models for weed identification in potato fields

Abstract

Weeds pose a significant and fundamental challenge in agriculture, competing with crops for vital resources such as water, nutrients, and sunlight. This competition often leads to reduced crop yields and diminished quality of produce. Additionally, weeds can host pests and diseases that further harm crops, increasing the risk of infestation and reducing farm productivity. Accurate weed identification through deep learning offers a solution, enabling farmers to implement site-specific herbicide spraying, thus lowering herbicide usage and minimizing environmental impact. This study introduces a benchmark crop and weed classification dataset and evaluates seven state-of-the-art deep-learning models for weed identification. The dataset was obtained from potato fields in Punjab, India, over two consecutive growth seasons (2022 and 2023). Seven deep learning models, Convolution Neural Network (CNN)-11, CNN-14, Inceptionv3, AlexNet, VGG16, ResNet50, and the YOLOv8 were trained and tested on this dataset for potato and weed classification. Among these models, YOLOv8 emerges as the top performer, achieving flawless accuracy of 100% with 37.5 million parameters. The custom CNN-11 model, despite having the fewest parameters (2.2 million), achieves 52% accuracy, making it suitable for resource-constrained environments. ResNet50, with its residual networks, also demonstrates exceptional performance with 99% accuracy and a moderate number of parameters (23 million), which can be a significant consideration in environments with limited resources or when deploying models on edge devices. These findings guide researchers and practitioners in selecting optimal models to reduce herbicide usage, minimize environmental impact, and enhance precision agriculture practices. Ultimately, this study advances weed management strategies, supporting sustainable crop management and improving agricultural productivity.