Table 2 Table 1 continued: Comparative overview of weed detection studies (Part 2).
References | Contribution | Dataset | Model Used | Results | Limitations |
|---|---|---|---|---|---|
Ukaegbu et al.40 | UAV-based sprayer with CNN for real-time weed detection | UAV images for weed classification | CNN-based model on Raspberry Pi | High accuracy in real-time detection | Battery and computational limitations |
Subeesh et al.45 | Detecting weeds in polyhouse-grown bell peppers (Capsicum annuum L.) using CNN | 1,106 images from a polyhouse | AlexNet, GoogLeNet, InceptionV3, Xception | InceptionV3 achieved 97.7% accuracy | Limited applicability to outdoor settings |
Dyrmann et al.29 | Classifying 22 plant species at early growth stages using CNN | 10,413 images from multiple sources | Custom CNN | 86.2% accuracy | High species similarity in early stages |
Wang et al.9 | Semantic segmentation for weed management with encoder-decoder network | Images of sugar beets (Beta vulgaris L. subsp. vulgaris var. altissima), oilseed rape (Brassica napus L. subsp. napus) | Encoder-decoder deep learning model | Highest MIoU of 88.91%, 96.12% accuracy | Dependent on NIR imagery |
Farooq et al.37 | Effect of spectral bands on weed classification with CNNs | Hyperspectral image dataset | CNN, compared with HoG | CNN with hyperspectral data achieved 97% accuracy | High-cost imagery required |
Arun et al.39 | Pixel-wise segmentation of crops/weeds using reduced U-Net | CWFID dataset | Reduced U-Net | 95% segmentation accuracy | Challenges in overlapping regions |
Olsen et al.28 | Developed DeepWeeds dataset for weed detection in rangeland environments | 17,509 images of 8 weed species from Australian rangelands | Inception-v3, ResNet-50 | ResNet-50 achieved 95.7% accuracy with 53.4 ms/image | Inter-class variability challenges |
Li and Zhang23 | Proposed DC-YOLO for crop and weed detection using YOLOv7-tiny | Public datasets and field-collected corn seedling data | DC-YOLO | mAP@0.5 of 95.7%; 5.223M parameters | Limited exploration of diverse weed types |
Sapkota et al.30 | Explored synthetic images for training Mask R-CNN | Real UAV images, synthetic images (real plant- and GAN-generated) | Mask R-CNN | Real plant-based synthetic images: mAPm of 0.60 | Synthetic images underperformed |
