An automated weed identification framework for sugarcane crop: A deep learning approach

Abstract

Automatic weed identification using deep learning (DL) models will mark a revolution in developing site– specific artificial intelligence (AI) based herbicide sprayers which intend to maximize herbicide efficiency with reduced herbicide application in agriculture production systems and hence contribute to higher yields. Conventional weed control strategies pose challenges for integrating smart herbicide delivery and machinery systems. This deep learning approach significantly impacts developing a system for weed identification required in establishing successful real time precision weed management systems like smart spraying systems and AI based smart machinery. Minimal research has been done on automatic weed identification in sugarcane (Saccharum officinarum L.) cropping systems. This study analyzed the feasibility of a computer vision based DL approach for weed identification to achieve autonomous weed control. The image dataset containing 5660 augmented images was deployed to train and evaluate DL models with a spilt of 90% for training and the rest for validation. We trained the six DL models for identifying weeds in sugarcane crop using field imagery (5094 images), validated (566 images) and further evaluated their accuracy and F1 score performance. Model training was undertaken by varying the hyperparameters, such as mini batch size (16 and 32) and epoch (10, 20 and 30) at a learning rate of 0.001. DarkNet53 accomplished a high F1 score value (>99%) and outperformed other models (AlexNet, GoogLeNet, InceptionV3, ResNet50, and Xception) for the identification of weeds in actively growing sugarcane crop. Weeds can be identified with a higher level of confidence (>98%) with a minimum error rate (<1%) at the mini batch size of 16 and epochs 20. Post-training and validation, DarkNet53 was tested (stage I) with an independent 200-image dataset followed by stage II testing (30 images) and obtained 96.6% net accuracy compared with naked eye weed identification. Based on the high – level performance of DarkNet53, we conclude that DL based weed identification laid a potential future avenue with an effective decision system in the machine vision aspects of a precision herbicide applicator for weed control in sugarcane fields through a low cost camera integrated with a single board computer.