András Barta, Gábor Horváth, Ákos Horváth, Ádám Egri, Miklós Blahó, Pál Barta, Karl Bumke, and Andreas Macke, "Testing a polarimetric cloud imager aboard research vessel Polarstern: comparison of color-based and polarimetric cloud detection algorithms," Appl. Opt. 54, 1065-1077 (2015)
Cloud cover estimation is an important part of routine meteorological observations. Cloudiness measurements are used in climate model evaluation, nowcasting solar radiation, parameterizing the fluctuations of sea surface insolation, and building energy transfer models of the atmosphere. Currently, the most widespread ground-based method to measure cloudiness is based on analyzing the unpolarized intensity and color distribution of the sky obtained by digital cameras. As a new approach, we propose that cloud detection can be aided by the additional use of skylight polarization measured by 180° field-of-view imaging polarimetry. In the fall of 2010, we tested such a novel polarimetric cloud detector aboard the research vessel Polarstern during expedition ANT-XXVII/1. One of our goals was to test the durability of the measurement hardware under the extreme conditions of a trans-Atlantic cruise. Here, we describe the instrument and compare the results of several different cloud detection algorithms, some conventional and some newly developed. We also discuss the weaknesses of our design and its possible improvements. The comparison with cloud detection algorithms developed for traditional nonpolarimetric full-sky imagers allowed us to evaluate the added value of polarimetric quantities. We found that (1) neural-network-based algorithms perform the best among the investigated schemes and (2) global information (the mean and variance of intensity), nonoptical information (e.g., sun-view geometry), and polarimetric information (e.g., the degree of polarization) improve the accuracy of cloud detection, albeit slightly.
You do not have subscription access to this journal. Citation lists with outbound citation links are available to subscribers only. You may subscribe either as an Optica member, or as an authorized user of your institution.
You do not have subscription access to this journal. Cited by links are available to subscribers only. You may subscribe either as an Optica member, or as an authorized user of your institution.
You do not have subscription access to this journal. Figure files are available to subscribers only. You may subscribe either as an Optica member, or as an authorized user of your institution.
You do not have subscription access to this journal. Article tables are available to subscribers only. You may subscribe either as an Optica member, or as an authorized user of your institution.
Five Filters Used in the Built-In Rotating Filter Wheel of Our Imaging Polarimetric Cloud Detectora
No.
Filter
1
Visible: infrared cut-off filter
2
Infrared: visible cut-off filter
3
Visible polarized: infrared cut-off and visible polarizer filters with 0° transmission angle
4
Visible polarized: infrared cut-off and visible polarizer filters with 45° transmission angle
5
Visible polarized: infrared cut-off and visible polarizer filters with 90° transmission angle
For calibration, we used a visible (optical) filter. Infrared measurements were strictly experimental.
Table 2.
Input Parameters Used by Our Four Neural-Network-Based Cloud Detection Algorithmsa
Input Parameter
PNN
NNN
SINN
Average in the red for the whole image
×
×
Average in the green for the whole image
×
×
Average in the blue for the whole image
×
×
Variance of in the red for the whole image
×
×
Variance of in the green for the whole image
×
×
Variance of in the blue for the whole image
×
×
Average in the red for the whole image
×
Average in the green for the whole image
×
Average in the blue for the whole image
×
Variance of in the red for the whole image
×
Variance of in the green for the whole image
×
Variance of in the blue for the whole image
×
Solar elevation
×
×
×
×
Average in the red for the given pixel
×
×
×
Average in the green for the given pixel
×
×
×
Average in the blue for the given pixel
×
×
×
Variance of in the red for the given pixel
×
×
Variance of in the green for the given pixel
×
×
Variance of in the blue for the given pixel
×
×
Average in the red for the given pixel
×
×
Average in the green for the given pixel
×
×
Average in the blue for the given pixel
×
×
Variance of in the red for the given pixel
×
Variance of in the green for the given pixel
×
Variance of in the blue for the given pixel
×
Elevation of the given pixel
×
×
×
×
Azimuth distance of sun and the given pixel
×
×
×
×
: intensity, : degree of linear polarization. Parameter usage in the PNN, NNN, SINN, and neural networks is marked with ×.
Table 3.
Fractional Error (%) for Each Algorithm of the Erroneously Classified Pixels Averaged over 10 Randomly Chosen Test Setsa
Algorithm (Detector)
Average Error (%)
Standard Deviation (%)
Using
Polarization Data
Global Parameters
Nonoptical Parameters
RBD
22.58
1.40
WD
21.46
1.45
RBR
21.64
1.41
19.63
1.14
HTA
20.78
1.61
×
WDAI
19.80
1.21
WDSD
18.57
1.00
×
42.88
2.69
×
29.94
2.35
×
23.82
1.15
×
×
SINN
16.32
1.12
×
NNN
16.05
1.24
×
×
PNN
15.32
1.09
×
×
×
RBD, red-blue difference; WD, whiteness detector; RBR, red-blue ratio; , -nearest neighbors; HTA, hybrid thresholding algorithm; WDAI, whiteness detector with average intensity; , degree of polarization difference in the green spectral range; , degree of polarization ratio in the green spectral range; WDSD, whiteness detector with solar distance; SINN, simple intensity neural network; , simple degree of polarization neural network; NNN, nonpolarimetric neural network; PNN, polarimetric neural network. The 10 sets of 25 test images were randomly chosen from the same pool of 50 images obtained during the ANT-XXVII/1 trans-Atlantic expedition of the research vessel Polarstern. The usage of polarization information, global parameters, and nonoptical parameters in the 13 different cloud detection algorithms is marked with ×.
Table 4.
Fractional Error (%) of the 13 Different Cloud Detection Algorithms Separately for the 10 Different Randomly Chosen Test Setsa
No.
RBD
WD
RBR
HTA
WDAI
WDSD
SINN
NNN
PNN
1
23.01
20.37
20.61
19.37
19.35
19.43
18.24
38.77
27.33
22.93
16.04
16.03
15.71
2
23.69
24.01
24.24
21.25
23.05
21.73
20.16
47.94
32.84
25.19
17.98
18.46
15.56
3
21.88
20.06
20.40
18.34
19.06
18.68
17.92
41.51
27.71
22.43
15.21
14.95
13.47
4
24.05
22.76
22.93
20.76
22.56
21.05
19.39
43.64
32.22
24.61
17.89
16.85
16.52
5
20.89
21.85
22.02
19.38
20.94
19.65
18.28
44.20
29.48
23.26
15.77
15.40
15.54
6
24.95
22.36
22.31
21.06
21.94
21.12
19.96
42.14
30.69
25.26
17.51
17.19
16.42
7
22.29
19.39
19.58
18.46
18.39
18.60
17.95
39.95
28.14
22.79
15.70
14.34
15.01
8
21.55
20.30
20.58
18.18
20.19
18.68
17.22
41.88
27.79
23.24
15.84
15.05
14.23
9
20.59
21.11
21.18
19.28
20.08
18.59
17.58
43.09
29.31
23.15
14.79
15.64
14.13
10
22.92
22.35
22.54
20.16
22.19
20.52
18.95
45.73
33.86
25.38
16.46
16.59
16.59
AV
22.58
21.46
21.64
19.63
20.78
19.80
18.57
42.88
29.94
23.82
16.32
16.05
15.32
SD
1.40
1.45
1.41
1.14
1.61
1.21
1.00
2.69
2.35
1.15
1.12
1.24
1.09
The average (AV) and standard deviation (SD) of the errors over the 10 test cases shown in the last two rows are from Table 3. RBD, red-blue difference; WD, whiteness detector; RBR, red-blue ratio; , -nearest neighbors; HTA, hybrid thresholding algorithm; WDAI, whiteness detector with average intensity; , degree of polarization difference in the green spectral range; , degree of polarization ratio in the green spectral range; WDSD, whiteness detector with solar distance; SINN, simple intensity neural network; , simple degree of polarization neural network; NNN, nonpolarimetric neural network; PNN, polarimetric neural network. For each test set as well as for the average, the smallest fractional error of the best algorithm is bolded and italicized.
Tables (4)
Table 1.
Five Filters Used in the Built-In Rotating Filter Wheel of Our Imaging Polarimetric Cloud Detectora
No.
Filter
1
Visible: infrared cut-off filter
2
Infrared: visible cut-off filter
3
Visible polarized: infrared cut-off and visible polarizer filters with 0° transmission angle
4
Visible polarized: infrared cut-off and visible polarizer filters with 45° transmission angle
5
Visible polarized: infrared cut-off and visible polarizer filters with 90° transmission angle
For calibration, we used a visible (optical) filter. Infrared measurements were strictly experimental.
Table 2.
Input Parameters Used by Our Four Neural-Network-Based Cloud Detection Algorithmsa
Input Parameter
PNN
NNN
SINN
Average in the red for the whole image
×
×
Average in the green for the whole image
×
×
Average in the blue for the whole image
×
×
Variance of in the red for the whole image
×
×
Variance of in the green for the whole image
×
×
Variance of in the blue for the whole image
×
×
Average in the red for the whole image
×
Average in the green for the whole image
×
Average in the blue for the whole image
×
Variance of in the red for the whole image
×
Variance of in the green for the whole image
×
Variance of in the blue for the whole image
×
Solar elevation
×
×
×
×
Average in the red for the given pixel
×
×
×
Average in the green for the given pixel
×
×
×
Average in the blue for the given pixel
×
×
×
Variance of in the red for the given pixel
×
×
Variance of in the green for the given pixel
×
×
Variance of in the blue for the given pixel
×
×
Average in the red for the given pixel
×
×
Average in the green for the given pixel
×
×
Average in the blue for the given pixel
×
×
Variance of in the red for the given pixel
×
Variance of in the green for the given pixel
×
Variance of in the blue for the given pixel
×
Elevation of the given pixel
×
×
×
×
Azimuth distance of sun and the given pixel
×
×
×
×
: intensity, : degree of linear polarization. Parameter usage in the PNN, NNN, SINN, and neural networks is marked with ×.
Table 3.
Fractional Error (%) for Each Algorithm of the Erroneously Classified Pixels Averaged over 10 Randomly Chosen Test Setsa
Algorithm (Detector)
Average Error (%)
Standard Deviation (%)
Using
Polarization Data
Global Parameters
Nonoptical Parameters
RBD
22.58
1.40
WD
21.46
1.45
RBR
21.64
1.41
19.63
1.14
HTA
20.78
1.61
×
WDAI
19.80
1.21
WDSD
18.57
1.00
×
42.88
2.69
×
29.94
2.35
×
23.82
1.15
×
×
SINN
16.32
1.12
×
NNN
16.05
1.24
×
×
PNN
15.32
1.09
×
×
×
RBD, red-blue difference; WD, whiteness detector; RBR, red-blue ratio; , -nearest neighbors; HTA, hybrid thresholding algorithm; WDAI, whiteness detector with average intensity; , degree of polarization difference in the green spectral range; , degree of polarization ratio in the green spectral range; WDSD, whiteness detector with solar distance; SINN, simple intensity neural network; , simple degree of polarization neural network; NNN, nonpolarimetric neural network; PNN, polarimetric neural network. The 10 sets of 25 test images were randomly chosen from the same pool of 50 images obtained during the ANT-XXVII/1 trans-Atlantic expedition of the research vessel Polarstern. The usage of polarization information, global parameters, and nonoptical parameters in the 13 different cloud detection algorithms is marked with ×.
Table 4.
Fractional Error (%) of the 13 Different Cloud Detection Algorithms Separately for the 10 Different Randomly Chosen Test Setsa
No.
RBD
WD
RBR
HTA
WDAI
WDSD
SINN
NNN
PNN
1
23.01
20.37
20.61
19.37
19.35
19.43
18.24
38.77
27.33
22.93
16.04
16.03
15.71
2
23.69
24.01
24.24
21.25
23.05
21.73
20.16
47.94
32.84
25.19
17.98
18.46
15.56
3
21.88
20.06
20.40
18.34
19.06
18.68
17.92
41.51
27.71
22.43
15.21
14.95
13.47
4
24.05
22.76
22.93
20.76
22.56
21.05
19.39
43.64
32.22
24.61
17.89
16.85
16.52
5
20.89
21.85
22.02
19.38
20.94
19.65
18.28
44.20
29.48
23.26
15.77
15.40
15.54
6
24.95
22.36
22.31
21.06
21.94
21.12
19.96
42.14
30.69
25.26
17.51
17.19
16.42
7
22.29
19.39
19.58
18.46
18.39
18.60
17.95
39.95
28.14
22.79
15.70
14.34
15.01
8
21.55
20.30
20.58
18.18
20.19
18.68
17.22
41.88
27.79
23.24
15.84
15.05
14.23
9
20.59
21.11
21.18
19.28
20.08
18.59
17.58
43.09
29.31
23.15
14.79
15.64
14.13
10
22.92
22.35
22.54
20.16
22.19
20.52
18.95
45.73
33.86
25.38
16.46
16.59
16.59
AV
22.58
21.46
21.64
19.63
20.78
19.80
18.57
42.88
29.94
23.82
16.32
16.05
15.32
SD
1.40
1.45
1.41
1.14
1.61
1.21
1.00
2.69
2.35
1.15
1.12
1.24
1.09
The average (AV) and standard deviation (SD) of the errors over the 10 test cases shown in the last two rows are from Table 3. RBD, red-blue difference; WD, whiteness detector; RBR, red-blue ratio; , -nearest neighbors; HTA, hybrid thresholding algorithm; WDAI, whiteness detector with average intensity; , degree of polarization difference in the green spectral range; , degree of polarization ratio in the green spectral range; WDSD, whiteness detector with solar distance; SINN, simple intensity neural network; , simple degree of polarization neural network; NNN, nonpolarimetric neural network; PNN, polarimetric neural network. For each test set as well as for the average, the smallest fractional error of the best algorithm is bolded and italicized.