Laser & Optoelectronics Progress, Volume. 59, Issue 18, 1810015(2022)
Cloud-Type Recognition Based on Multiscale Features and Gradient Information
Figures & Tables(16)
Fig. 4. Comparison between DGNet121 and classical classification networks. (a) Training loss; (b) testing accuracy
Fig. 6. Classification accuracy of DGNet network with different position gradient features
Table 1. Characteristics and number of each cloud type
View tableTable 1. Characteristics and number of each cloud type
Cloud type Number of collected images Number of experimental images Characteristic description Cirrus 1293 100 Relatively thin，pinnate，white filamentous Cirrocumulus 391 100 Slight white wavy，scaly or globose cloud Stratocirrus 317 100 White transparent filamentous structure Cumulus 1889 100 The individual is noticeable，flat at the bottom，the light part is white，and the bottom is dark Cumulonimbus 886 100 The clouds are thick and big，dark and messy Stratocumulus 803 100 Clouds are large，in strips，sheets，or lumps Nimbostratus 6681 100 The clouds form an even curtain，shielding the sun and moon Altocumulus 359 100 Clouds are small，usually oval，tile，or fish scale-shaped Altostratus 2498 100 Clouds are grayish-white or gray with striated structures at the base Stratus 646 100 Gray foggy
Table 2. Comparison of classical classification frameworks
View tableTable 2. Comparison of classical classification frameworks
Method Top-1 /% Top-5 /% Inception_v1 54.00 90.33 Inception_v2 54.30 91.67 Inception_v3 60.30 95.67 Inception_v4 50.70 90.33 ResNet50 63.70 95.33 ResNet101 61.00 94.33 DenseNet121 64.00 97.33 DGNet121 64.30 96.00
Table 3. Identification accuracy of classical classification frameworks for each cloud type
View tableTable 3. Identification accuracy of classical classification frameworks for each cloud type
Method Stratocumulus Stratus Altostratus Altocumulus Cumulus Cumulonimbus Stratocirrus Cirrocumulus Cirrus Nimbostratus Inception_v1 46.67 86.67 3.33 73.33 46.67 80 60 53.33 46.67 43.33 Inception_v2 40.00 86.67 30.00 83.33 56.67 80 63.33 23.33 60.00 20.00 Inception_v3 43.33 90.00 33.33 76.67 53.33 70.00 80.00 73.33 53.33 30.00 Inception_v4 26.67 76.67 6.67 73.33 56.67 76.67 56.67 43.33 50.00 40.00 ResNet50 33.33 76.67 46.67 86.67 76.67 73.33 60.00 60.00 56.67 66.67 ResNet101 46.67 66.67 40.00 86.67 66.67 66.67 66.67 63.33 63.33 43.33 DenseNet121 53.33 76.67 36.67 83.33 70.00 63.33 66.67 73.33 60.00 56.67 DGNet121 60.00 73.33 33.33 83.33 70.00 76.66 70.00 76.67 60.00 40.00
Table 4. Comparison of different loss calculations for DGNet121
View tableTable 4. Comparison of different loss calculations for DGNet121
Parameter 1 2 3 4 5 Mean Variance LDL1 /% 64.30 64.30 63.70 69.70 63.30 65.06 0.06908 LDL2 /% 66.30 65.00 67.30 64.30 63.00 65.18 0.02827
Table 5. Comparison experiment of gradient features on DGNet121 network
View tableTable 5. Comparison experiment of gradient features on DGNet121 network
Method Top-1 /% Top-5 /% DGNet121_1+2 65.70 97.67 DGNet121_1+3 62.00 94.33 DGNet121_1+4 67.00 97.00 DGNet121_1+2+3 64.70 95.00 DGNet121_1+2+4 63.30 98.33 DGNet121_1+2+3+4 63.00 96.67
Table 6. Comparison experiment of different proportion training and test sets on DGNet121 network
View tableTable 6. Comparison experiment of different proportion training and test sets on DGNet121 network
Item Ntraining∶Ntesting Accuracy /% A 70∶30 64.00 B 150∶30 69.30 C 150∶50 74.40 D 200∶50 64.70 E 200∶100 67.90
Tools
Get Citation
Copy Citation Text
Lingjie Jin, Zhiwei Lin, Yu Hong. Cloud-Type Recognition Based on Multiscale Features and Gradient Information[J]. Laser & Optoelectronics Progress, 2022, 59(18): 1810015
Paper Information
Category: Image Processing
Received: Jul. 8, 2021
Accepted: Aug. 10, 2021
Published Online: Aug. 29, 2022
The Author Email: Lin Zhiwei (cwlin@fafu.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20