Fig. 1. Model architecture diagram
Fig. 2. Original labeled data. (a) Sample 1; (b) sample 2; (c) sample 3; (d) sample 4
Fig. 3. Fused data
Fig. 4. Satellite cloud image data. (a) Sample 1; (b) sample 2; (c) sample 3; (d) sample 4
Fig. 5. Sample drawing of detect results. (a) Sample 1; (b) sample 2; (c) sample 3; (d) sample 4
Fig. 6. Loss function figure of Mask R-CNN model
Fig. 7. Loss function figures of Mask R-CNN model combined with data augmentation. (a) Proposed multi-scale mosaic; (b) Cutout; (c) CutMix; (d) mosaic
Fig. 8. Deep learning method for locating typhoon center. (a) Faster R-CNN; (b) YOLOv3; (c) Mask R-CNN
Fig. 9. Fitting diagrams of real coordinates and segmented coordinates of model. (a) HAISHEN; (b) VAMCO
Level of scale | Radius size | Number of pictures |
---|
Small | < 1.43 | 1467 | Middle | 1.43 ≤< 1.90 | 1467 | Large | ≥ 1.90 | 1466 |
|
Table 1. Scale division of typhoon eye
Hyperparameter | Value |
---|
NMS-thresh | 0.45 | Score-thresh | 0.5 | Maximum iterations | 40000 | Model checkpoint | 2 | Learning rate | 10-4-10-6 | Weight decay | 5×10-4 | GPU number | 1 | Input size | 128×128 |
|
Table 2. Universal hyperparameter settings
Scale combination | Number | MSE | Mean MSE | MIoU | Mean MIoU | MPA | Mean MPA | FPS | Mean FPS |
---|
Master model | 1 | 3802 | | 0.9086 | | 0.8959 | | 20.4 | | Large+Middle | 1 | 2154 | 2554 | 0.9454 | 0.9421 | 0.9263 | 0.9218 | 17.1 | 20.6 | 2 | 2981 | 0.9384 | 0.9170 | 23.3 | 3 | 2527 | 0.9426 | 0.9221 | 21.5 | Large+Small | 1 | 3290 | 3415 | 0.9322 | 0.9313 | 0.9098 | 0.9115 | 19.4 | 21.4 | 2 | 3747 | 0.9270 | 0.9061 | 22.7 | 3 | 3208 | 0.9348 | 0.9186 | 22.2 | Middle+Small | 1 | 3685 | 4176 | 0.9259 | 0.9151 | 0.8836 | 0.8896 | 20.8 | 20.7 | 2 | 4761 | 0.9151 | 0.8908 | 19.8 | 3 | 4082 | 0.9044 | 0.8944 | 21.5 | Large+Middle+Small | 1 | 3125 | 3355 | 0.9187 | 0.9222 | 0.8968 | 0.9043 | 22.1 | 20 | 2 | 3588 | 0.9221 | 0.9047 | 18.7 | 3 | 3351 | 0.9258 | 0.9113 | 19.2 |
|
Table 3. Results of multi-scale experiments
Model | MIoUS | MIoUM | MIoUL | MPAS | MPAM | MPAL |
---|
Mask R-CNN | 0.8880 | 0.9145 | 0.9233 | 0.8830 | 0.8831 | 0.9216 | Mask R-CNN+Proposed multi-scale mosaic | 0.9305 | 0.9469 | 0.9489 | 0.9090 | 0.9197 | 0.9367 |
|
Table 4. Comparison of wind eye segmentation results at different scales
Model | Data augmentation method | MSE | MIoU | MPA | FPS |
---|
Mask R-CNN | | 3802 | 0.9086 | 0.8959 | 20.4 | Proposed multi-scale mosaic | 2554 | 0.9421 | 0.9218 | 20.6 | Cutout | 5373 | 0.8569 | 0.8247 | 22.8 | CutMix | 4006 | 0.9135 | 0.9061 | 17.0 | mosaic | 3249 | 0.9207 | 0.9074 | 15.4 |
|
Table 5. Experimental results of each data augmentation
Model | Mean error | Mean variance |
---|
Longitude | Latitude | Longitude | Latitude |
---|
Faster R-CNN | 0.31 | 0.26 | 0.055 | 0.046 | YOLOv3 | 0.28 | 0.25 | 0.043 | 0.039 | Mask R-CNN+Proposed multi-scale mosaic | 0.17 | 0.19 | 0.028 | 0.025 |
|
Table 6. Comparison of location accuracy of typhoon center