The image recognition algorithm used for SCASI cloud detection is a variation of shape analysis used in soil image recognition systems. The first step is to create a mask so that only the sky part of the image is included in the analysis, trees, mountains, the telescope dome and the secondary mirror must all be ignored. Thresholding image segmentation is used to generate a binary block object image. Object analysis is performed on the binary block image, analysing the number of objects, the size of each object and the percentage of the image covered by objects.
This data is then compared to a statistical empirically derived table to decide which category the image falls in. Images that are borderline or do not conform to the empirical data are flagged as unidentified and require manual identification. The empirical look-up table was derived from just over 3000 SCASI images, manually identified. This was checked against 7000 SCASI images, where,
- 90.1% of images were identified correctly
- 9.9% were unidentified
- 0% were misidentified.
SCASI cloud detection plays an important role in the creation of characteristic energy maps, as it is vital that the true intensity of the aurora is measured. The cloud detection dataset is also included in the STP events database so that searches may be restricted to periods of clear skies at Skibotn.