Introduction
Materials and Methods
Study Area and Data
Insect Annotation
Data Augmentation
Year | Augment tech | Total added images |
---|---|---|
2020 | Flippinga | 806 |
2021 | Flippinga | 1203 |
2021 | Grid croppingb | 50,110 |
2021 | Sliding window croppingb | 177,325 |
Deep Learning Models
Backbones and Pre-training
Classification Model Evaluation Metrics
Ground truth class/Predicted class | RHAGCE | Non-RHAGCE |
---|---|---|
RHAGCE | Count of TPs | Count of FNs |
Non-RHAGCE | Count of FPs | Count of TNs |
Results and Discussions
Insect Annotation Results
2020 annotation (140 Images—47 classes) | |||
---|---|---|---|
Scientific name | Class-Name | Insect count | Observed in # images |
Indefinable insects | NOTHIN | 2644 | 89 |
Background of image | BACKGR | 885 | 23 |
Rhagoletis cerasi | RHAGCE | 469 | 84 |
Drosophila | 1DROSG | 183 | 38 |
Chrysopidae | 1CHASF | 151 | 12 |
Muscidae | 1MUSCF | 109 | 31 |
Formicidae | 1CHSAF | 100 | 19 |
Other | 40-Classes | 364 | 114 |
2021 annotation (850 Images—1 class) | |||
Rhagoletis cerasi | RHAGCE | 1626 | 401 |
2020 annotation | |||
---|---|---|---|
Scientific name | Class-Name | Insect count | Observed in # images |
Indefinable insects | NOTHIN | 16,638 | 610 |
Background of image | BACKGR | 6048 | 164 |
Rhagoletis cerasi | RHAGCE | 3424 | 573 |
Drosophila | 1DROSG | 1073 | 216 |
Chrysopidae | 1CHASF | 919 | 78 |
Muscidae | 1MUSCF | 775 | 215 |
Formicidae | 1CHSAF | 393 | 97 |
2021 annotation | |||
Rhagoletis cerasi | RHAGCE | 4872 | 1203 |
Deep Learning Models Evaluation
DL model | Learning rate (lr) | AP@0.5 | AP@0.75 | Average fps |
---|---|---|---|---|
Faster R‑CNN MobileNet | 0.01 | 0.88 | 0.69 | 15.47 |
Faster R‑CNN MobileNet | 0.001 | 0.88 | 0.59 | 15.76 |
RetinaNet ResNet | 0.001 | 0.88 | 0.6 | 11.91 |
SSD VGG-16 | 0.0001 | 0.88 | 0.55 | 9.33 |
SSD VGG-16 | 0.001 | 0.87 | 0.57 | 9.48 |
Faster R‑CNN ResNet | 0.001 | 0.86 | 0.59 | 10.99 |
RetinaNet ResNet | 0.0001 | 0.86 | 0.42 | 11.88 |
Faster R‑CNN ResNet | 0.0001 | 0.84 | 0.53 | 9.83 |
SSD VGG-16 | 0.00001 | 0.84 | 0.35 | 9.33 |
SSD VGG-16 no pretrain | 0.0001 | 0.83 | 0.37 | 9.33 |
Faster R‑CNN MobileNet | 0.0001 | 0.81 | 0.33 | 15.63 |
YOLOV5 | 0.001 | 0.76 | 0.73 | 15.04 |
YOLOV5 | 0.0001 | 0.76 | 0.67 | 14.75 |
YOLOV5 | 0.01 | 0.75 | 0.75 | 14.21 |
SSD VGG-16 no pretrain | 0.00001 | 0.75 | 0.18 | 9.24 |
Faster R‑CNN ResNet | 0.01 | 0.51 | 0.71 | 10.96 |
RetinaNet ResNet | 0.00001 | 0.28 | 0.1 | 10.34 |
Faster R‑CNN MobileNet | 0.00001 | 0.19 | 0.09 | 13.35 |
YOLOV5 | 0.00001 | 0 | 0 | 14.87 |
DL model | Learning rate (lr) | mAP@0.5 | Average fps |
---|---|---|---|
Faster R‑CNN ResNet | 0.01 | 0.51 | 1.44 |
Faster R‑CNN ResNet | 0.0001 | 0.14 | 1.72 |
Faster R‑CNN MobileNet | 0.001 | 0.19 | 1.52 |
Faster R‑CNN MobileNet | 0.0001 | 0.1 | 1.37 |
RetinaNet ResNet | 0.001 | 0.42 | 1.2 |
RetinaNet ResNet | 0.00001 | 0 | 1.74 |