Table 5.
A summary of new trends for the orchard-UAV-NN triplet.
| Model Novelty | Characteristics, Pros, and Cons | NN used and function | Performance indicators | References |
|---|---|---|---|---|
| ▪Combining two different CNNs |
▪Semantic segmentation of vegetation. ▪Pros: Good results in a wetland mapping application. ▪Cons: Slower training process. |
▪SegNet with VGG16 ▪SegNet with ResNet50 ▪UNet with VGG16 ▪UNet with ResNet50 |
▪ACC = 91% for SegNet with ResNet50 ▪Time for NN training: 700 min |
(Bhatnagar et al., 2020) |
| ▪Fusing the outputs of two CNN, one for RGB and the other for NIR images |
▪Two camera sensors for RGB and NIR. Disease detection in vine crops using segmentation ▪Pros: Fusion by intersection is better than classes detected in the visible or infrared range: ▪Cons: Reduced performances on segmentation due to the small training set and too few NNs in the system, long runtime |
▪Two SegNet (RGB and NIR) ▪Two LeNet5 (RGB and NIR) for pre-labeling |
▪Leaf-level average ACC: 82.20% - fusion AND; 90.23% - fusion OR; ▪Grapevine-level average ACC: 88.14% - fusion AND; 95.02% - fusion OR; |
(Kerkech et al., 2020) |
| ▪Net with a specific name for the application: DeepSolanum- |
▪Segmentation of UAV images to detect the invasion degree of “Solanum rostratum Dunal” ▪Pros: Reduced training time and complexity ▪Cons: Performances must be improved |
▪DeepSolanum-Net based on U-Net |
▪Precision = 89.95% ▪Recall = 90.3% ▪IoU = 82.76% ▪F1-score = 89.85% |
(Wang et al., 2021) |
|
▪Different CNN combined in a system for orchard monitoring ▪Net with a specific name: MangoYOLO |
▪Detect and count the fruits within images. Input: tree image. Output: total fruits per tree ▪Pros: Good performance for fruit counting in one season. ▪Cons: It is not a robust model in different seasons. |
▪Multi Layered Perceptron (MLP), ▪MangoYOLO model, ▪Xception_count model with a regression block, ▪Xception_classification model |
▪Best R2 = 94% | (Koirala et al., 2021) |
| ▪Including a CNN as a backbone in other CNN |
▪Detection and semantic segmentation of coconut trees ▪Pros: Good ACC ▪Cons: Need to classify and locate different kinds of trees. |
▪Mask R-CNN with ResNet 101 as a backbone |
▪mAP = 91% ▪ACC (classification) = 97% |
(Iqbal et al., 2021) |
| ▪Dual network-based system to eliminate successively some FN and FP errors |
▪Detecting and classifying harmful insects in orchards (HH) ▪ Pros: Good performance to detect insects in the foreground. ▪ Cons: Need to detect insects in a distant plane. |
▪YOLOv.4 with DarkNet combined with EfficientNet B3 |
▪ACC = 95% ▪F1-score = 92% |
(Popescu et al., 2022b) |
| ▪Combining NN YOLOv5s, DeepLabv3+ MobileNetv2 |
▪Detecting and segmentation of the logan fruit branch for logan harvesting using RGB-D camera ▪Pros: Reduced operating time and good ACC semantic segmentation ▪Cons: Limitations of object detection and segmentation in environmental interference conditions |
▪Improved YOLOv5s for detection and DeepLabv3+ MobileNetv2 for semantic segmentation |
▪ACC = 85.50% (fruit branch detection) ▪ACC = 94.52% (fruit branch semantic segmentation) |
(Li, D. et al., 2022) |
| ▪Faster R-CNN improved with the Feature Pyramid Networks (FPN) |
▪Count the number of pecans in an orchard ▪Pros: Good mAP to identify pecans ▪Cons: Influence of lighting on fruit recognition and detection. |
▪Faster R-CNN and FPN | ▪mAP = 95.932% | (Hu et al., 2022) |
| ▪Federated learning (FL) and improved Faster R-CNN. |
▪Multiple pest detection ▪Pros: Can detect multiple pests in a short time. ▪Cons: ACC must be improved |
▪Faster RCNN with ResNet 101 and with FL |
▪mAP = 89.34% ▪ACC = 90.27% ▪Detection time = 0.05 s |
(Deng et al., 2022) |
| ▪Combining three improved DensNet 121 |
▪Pest detection from an augmented big dataset ▪Pros: Detecting pests on various agricultural crops ▪Cons: Performances must be improved |
▪Improved three DensNet 121 and combined them into a decision fusion system | ▪ACC = 75.28% | (Peng et al., 2023) |