Optimized Weed Image Classification via Parallel Convolutional Neural Networks Integrating an Excess Green Index Channel

Weed management is an essential operational task to ensure the excellent health of crops or trees. The emergence of machine vision enables convolutional neural networks (CNNs) to classify weed types automatically, which can subsequently be used for a weed management strategy. A dominant approach to implement CNN-based weed classification is to train a network with RGB images as input either by adopting a transfer learning approach or a custom network. However, such an approach limits the process of incorporating prior knowledge as a significant feature of the network to improve the classification accuracy. This work proposes a novel network based on parallel convolutional neural networks (P-CNN), leveraging the excess green index (ExG) channel as an additional input to the RGB image channels. We argue that using the ExG channel can capture the greenness feature of weeds from the visible light spectrum, an important feature in many vegetation images such as leaves or green plants. The results show that the proposed P-CNN combining ResNet50 and a custom CNN obtains a Top-1 accuracy of 97.2% on a public weed dataset called DeepWeeds compared to the baseline ResNet50 alone with only 95.7%. The results show the significant contribution of domain- specific knowledge of green indexes in improving the classification performance of weed images. This enhancement could transform real-world weed management by enabling highly precise detection by allowing the classifier to focus intensively on differentiating green color features between leaves with nearly identical morphology.