Table 1 Comparative overview of weed detection studies (Part 1): Analysis of contributions, datasets, models, results, and limitations in weed detection research.
References | Contribution | Dataset | Model used | Results | Limitations |
|---|---|---|---|---|---|
Hussain et al.21 | Detecting common lambsquarters (Chenopodium album L.) in potato (Solanum tuberosum L.) fields; used CNN with PHA image encoding | 30,160 images from Atlantic Canada potato field | GoogLeNet, VGG-16, EfficientNet | EfficientNet achieved 92-97% accuracy, outperforming other models | Requires sophisticated imaging equipment and high computational resources |
Islam et al.35 | UAV-based early weed detection in chilli pepper (Capsicum annuum L.) farms | UAV imagery from Australian chilli pepper fields | Random Forest, SVM, KNN | RF achieved highest accuracy at 96%; SVM 94%, KNN 63% | Limited by UAV flight conditions and image resolution |
Xu et al.33 | Multi-modal deep learning with RGB-D for weed detection in wheat (Triticum aestivum L.) crop | Wheat field images with RGB-D modality | Custom three-channel network for RGB-D | 89.3% precision (IoG) | Requires RGB-D cameras and high computational resources |
Almalky and Ahmed11 | Drone-based detection and classification of Consolida regalis L. weed growth stages using deep learning models | 3731 images of Consolida regalis weed at four growth stages | YOLOv5, RetinaNet, Faster R-CNN | YOLOv5-small achieved recall of 0.794; RetinaNet achieved AP of 87.457% | Relied on a single species dataset |
Teimouri et al.12 | Estimating weed growth stages based on leaf counts using CNN | 9649 RGB images of 18 weed species across nine growth stages | Inception-v3 CNN architecture | Accuracy of 78% for Polygonum spp.; 70% overall accuracy | Challenges in overlapping leaves and inconsistent performance |
Costello et al.44 | Detection and growth stage classification of ragweed parthenium (Parthenium hysterophorus L.) using RGB and hyperspectral imagery | 665 RGB images and hyperspectral data in controlled environments | YOLOv4 for RGB; XGBoost for hyperspectral | YOLOv4: 95% detection, 86% classification; XGBoost: 99% classification | Limited applicability to field conditions |
Lottes et al.34 | Fully convolutional network using sequential data for crop-weed classification | RGB+NIR images from sugar beet fields | Fully Convolutional Network | Over 94% recall for crops, 91% for weeds | Needs consistent image sequences and high computational resources |
Peteinatos et al.22 | Identified 12 plant species using deep learning | 93,130 labeled images under field conditions | VGG16, ResNet-50, Xception | ResNet-50, Xception achieved >97% accuracy; VGG16 82% | Controlled imaging conditions needed |
Beeharry and Bassoo36 | Evaluated ANN and AlexNet for UAV weed detection | 15,336 segmented images | ANN, AlexNet | AlexNet achieved 99.8% accuracy; ANN 48.09% | Computational demands for UAV |
Jeon et al.38 | Adaptive algorithm for plant segmentation under variable lighting | 666 field images | ANN with image segmentation | 95.1% accuracy for crops | Limited scalability across environments |
