Applications of deep learning in precision weed management: A review

Abstract

Deep Learning (DL) has been described as one of the key subfields of Artificial Intelligence (AI) that is transforming weed detection for site-specific weed management (SSWM). In the last demi-decade, DL techniques have been integrated with ground as well as aerial-based technologies to identify weeds in still image context and real-time setting. After observing the current research trend in DL-based weed detection, techniques are advancing by assisting precision weeding technologies to make smart decisions. Therefore, the objective of this paper was to present a systematic review study that involves DL-based weed detection techniques and technologies available for SSWM. To accomplish this study, a comprehensive literature survey was performed that consists of 60 closest technical papers on DL-based weed detection. The key findings are summarized as follows, a) transfer learning approach is a widely adopted technique to address weed detection in majority of research work, b) less focus navigated towards custom designed neural networks for weed detection task, c) based on the pretrained models deployed on test dataset, no one specific model can be attributed to have achieved high accuracy on multiple field images pertaining to several research studies, d) inferencing DL models on resource-constrained edge devices with limited number of dataset is lagging, e) different versions of YOLO (mostly v3) is a widely adopted model for detecting weeds in real-time scenario, f) SegNet and U-Net models have been deployed to accomplish semantic segmentation task in multispectral aerial imagery, g) less number of open-source weed image dataset acquired using drones, h) lack of research in exploring optimization and generalization techniques for weed identification in aerial images, i) research in exploring ways to design models that consume less training hours, low-power consumption and less parameters during training or inferencing, and j) slow-moving advances in optimizing models based on domain adaptation approach. In conclusion, this review will help researchers, DL experts, weed scientists, farmers, and technology extension specialist to gain updates in the area of DL techniques and technologies available for SSWM.