/

Vessel Attribute Estimations

For vessels detected in Sentinel-2 Optical Imagery, Skylight additionally estimates length and heading for the vessel based on the image. This is particularly helpful for analyzing “dark” vessels where no additional information is available via AIS. 

Length Estimation

Here is the Confusion Matrix for our model that predicts vessel length. “Actual Category” is the length as reported in the AIS data correlated with the detection, while “Predicted Category” is the length the model determined from the image itself. Examples of how to interpret this matrix: 

  • 73% of vessels with actual length between 100-150m are predicted with some length value falling between 100-150m
  • 89% of the vessels predicted to be between 10–20 m are actually within one bucket on either side (0–30 m) 
   

Predicted Category

 
   

0-10

10-20

20-30

30-50

50-75

75-100

100-150

150-200

200+

Total
Actual Category

0-10

0

129 (9%)

218 (4%)

88 (1%)

39 (1%)

12 (0%)

7 (0%)

0 (0%)

2 (0%)

495

10-20

0

1058 (71%)

2632 (45%)

867 (12%)

221 (4%)

80 (2%)

71 (1%)

27 (0%)

41 (0%)

4997

20-30

0

137 (9%)

1817 (31%)

1944 (27%)

373 (6%)

169 (3%)

111 (1%)

64 (1%)

54 (0%)

4669

30-50

0

78 (5%)

695 (12%)

3222 (44%)

1419 (24%)

235 (5%)

167 (2%)

57 (1%)

62 (1%)

5935

50-75

0

28 (2%)

165 (3%)

602 (8%)

2905 (49%)

982 (20%)

150 (2%)

48 (0%)

32 (0%)

4912

75-100

0

14 (1%)

92 (2%)

166 (2%)

606 (10%)

2737 (56%)

1172 (15%)

68 (1%)

44 (0%)

4899

100-150

0

14 (1%)

92 (2%)

186 (3%)

171 (3%)

495 (10%)

5783 (73%)

852 (9%)

38 (0%)

7631

150-200

0

17 (1%)

35 (1%)

150 (2%)

79 (1%)

76 (2%)

385 (5%)

8432 (85%)

1514 (12%)

10733

200+

0

12 (1%)

21 (1%)

102 (1%)

73 (1%)

74 (2%)

73 (1%)

350 (4%)

10549 (86%)

11279
  Total 0 1487 5837 7327 5886 4860 7919 9898 12336 55550

Note: This model is regressing the length (predicting continuous numerical values) and the confusion matrix here is produced by putting the ground truth and predicted length into buckets. This may also explain part of why the model performance is slightly worse when looking at correct category, but better when looking one bucket off. Small (meter-scale) errors often push a prediction into the adjacent bin, so exact-bucket accuracy falls while “within one bucket” accuracy rises.

 

Heading Estimation

79% of heading estimates are within 10 degrees of the ground truth heading.

The arrow on the top right of vessel image chips indicates the direction of the predicted heading.