264
An Automatic Count of a Starling Murmuration at Aberystwyth,
Ceredigion, UK
One of the most difficult challenges for any bird counter is to assess the number of
Starlings participating in aerial displays (murmurating) before going to roost.
The intricate formations witnessed during Starling murmurations are spellbinding and can involve tens of thousands of birds. Reliably estimating the number of
birds involved in these displays by eye is impossible. However, if we can capture an
image of a flock of birds against a plain sky background, there is a good chance that
it will be suitable for an automatic count in ImageJ. In images with a plain background, the factor most likely to influence the accuracy of the count will be the
degree of overlap between individual birds. In images of dense flocks, there will
always be some individuals overlapping, and in dense flocks comprising a large
number of birds, this could equate to thousands. This will pose a problem for any
program that bases its counts on the number of objects in an image, because ten
overlapping birds could form a single object and contribute only a single bird to the
total. The impact of this problem can be minimised, however, because ImageJ provides two options for validating the counts: a drawing showing the numbered objects
that contributed to the count and an Excel file stating the pixel size of every object
that contributed to the count. In practice, the drawing is an excellent source of confidence that the objects counted by the program actually equate to the birds in the
original image. If this is not the case, then the settings for image threshold and
object pixel size will need adjusting. Text Box 2 describes the process involved in
carrying out the automatic count of the Starling murmuration shown in Fig. 3, while
Fig. 11 shows an inset from the drawing showing the numbered objects that were
counted, this indicated that all of the objects counted were birds.
Starling Case Study Results and Validation
The total generated by the automatic count in ImageJ was 9380 birds, but this
included a considerable number of overlapping birds. The validation process
described below improved the accuracy of this estimate.
Validation
The automatic count details in the Excel table (Table 2) suggested that the area of
most individuals in the Starling flock was in the 35–80 pixel range, depending on
the position of the bird in the flock (closer or further from the camera) and whether
the wings were outstretched or close to the body at the time. However, because there
C. Hurford
An Automatic Count of a Starling Murmuration at Aberystwyth,
Ceredigion, UK
One of the most difficult challenges for any bird counter is to assess the number of
Starlings participating in aerial displays (murmurating) before going to roost.
The intricate formations witnessed during Starling murmurations are spellbinding and can involve tens of thousands of birds. Reliably estimating the number of
birds involved in these displays by eye is impossible. However, if we can capture an
image of a flock of birds against a plain sky background, there is a good chance that
it will be suitable for an automatic count in ImageJ. In images with a plain background, the factor most likely to influence the accuracy of the count will be the
degree of overlap between individual birds. In images of dense flocks, there will
always be some individuals overlapping, and in dense flocks comprising a large
number of birds, this could equate to thousands. This will pose a problem for any
program that bases its counts on the number of objects in an image, because ten
overlapping birds could form a single object and contribute only a single bird to the
total. The impact of this problem can be minimised, however, because ImageJ provides two options for validating the counts: a drawing showing the numbered objects
that contributed to the count and an Excel file stating the pixel size of every object
that contributed to the count. In practice, the drawing is an excellent source of confidence that the objects counted by the program actually equate to the birds in the
original image. If this is not the case, then the settings for image threshold and
object pixel size will need adjusting. Text Box 2 describes the process involved in
carrying out the automatic count of the Starling murmuration shown in Fig. 3, while
Fig. 11 shows an inset from the drawing showing the numbered objects that were
counted, this indicated that all of the objects counted were birds.
Starling Case Study Results and Validation
The total generated by the automatic count in ImageJ was 9380 birds, but this
included a considerable number of overlapping birds. The validation process
described below improved the accuracy of this estimate.
Validation
The automatic count details in the Excel table (Table 2) suggested that the area of
most individuals in the Starling flock was in the 35–80 pixel range, depending on
the position of the bird in the flock (closer or further from the camera) and whether
the wings were outstretched or close to the body at the time. However, because there
C. Hurford