274
The total generated by the automatic count was 832: this was revised to 848 birds
after using the pixel sizes in the Excel file to validate the count. A manual count
revealed 842 birds in the image. Therefore, the automatic count of 832 was an
underestimate of almost 1% and the validation of the automatic count of 848 was an
overestimate of c.0.7%.
Discussion
The results of the exercises described in this chapter suggest that the range of
observer variation associated with ‘by eye’ bird counts is considerably greater than
any real change we would tolerate without some form of political or conservation
management response. The results of these exercises should be a concern for any
ornithologist or researcher using ‘by eye’ count estimates for scientific purposes. In
contrast, we can expect the results of careful manual counts in ImageJ to be 99–100%
accurate, with the accuracy of automatic counts also likely to exceed 90% accuracy,
particularly after being revised through validation.
The results from the case studies suggest that technology has a significant role to
play in improving the accuracy of bird counts, particularly on breeding grounds, in
wintering areas and at staging sites during migration periods. This is particularly
true where counts involve gregarious species which, in western Europe, include
Common Scoters (Melanitta nigra), Cranes (Grus grus), Whooper Swans (Cygnus
cygnus), White Storks (Ciconia ciconia), Greater Flamingos, Guillemots, Golden
Plovers, Lapwings, Dunlins (Calidris alpina), Knots (Calidris canutus), Starlings,
and many other species of geese, wader, wildfowl, lark, pipit, finch and bunting.
There are a small number of examples in the literature of researchers experimenting with remote counts using segmentation analysis methods (e.g. Chabot and
Francis 2016, Merkel et al. 2016, Delord et al. 2015, and Groom et al. 2013) but the
recommendations have not yet penetrated mainstream thinking in ornithological
circles. Some of these segmentation exercises have been carried out in eCognition
software (Nussbaum and Menz 2008), but this software is both specialised and
expensive and would not be readily available or accessible to most birdwatchers, in
direct contrast to ImageJ software.
Given suitable images to process, ImageJ can carry out relatively accurate automatic counts of large aggregations of birds. These automatic counts take seconds to
complete and come with three options for validation, (a) a numbered drawing of the
objects contributing to the count, (b) an Excel file listing the pixel sizes of each object
in the drawing and (c) partial or complete manual counts of the birds in the image.
C. Hurford
The total generated by the automatic count was 832: this was revised to 848 birds
after using the pixel sizes in the Excel file to validate the count. A manual count
revealed 842 birds in the image. Therefore, the automatic count of 832 was an
underestimate of almost 1% and the validation of the automatic count of 848 was an
overestimate of c.0.7%.
Discussion
The results of the exercises described in this chapter suggest that the range of
observer variation associated with ‘by eye’ bird counts is considerably greater than
any real change we would tolerate without some form of political or conservation
management response. The results of these exercises should be a concern for any
ornithologist or researcher using ‘by eye’ count estimates for scientific purposes. In
contrast, we can expect the results of careful manual counts in ImageJ to be 99–100%
accurate, with the accuracy of automatic counts also likely to exceed 90% accuracy,
particularly after being revised through validation.
The results from the case studies suggest that technology has a significant role to
play in improving the accuracy of bird counts, particularly on breeding grounds, in
wintering areas and at staging sites during migration periods. This is particularly
true where counts involve gregarious species which, in western Europe, include
Common Scoters (Melanitta nigra), Cranes (Grus grus), Whooper Swans (Cygnus
cygnus), White Storks (Ciconia ciconia), Greater Flamingos, Guillemots, Golden
Plovers, Lapwings, Dunlins (Calidris alpina), Knots (Calidris canutus), Starlings,
and many other species of geese, wader, wildfowl, lark, pipit, finch and bunting.
There are a small number of examples in the literature of researchers experimenting with remote counts using segmentation analysis methods (e.g. Chabot and
Francis 2016, Merkel et al. 2016, Delord et al. 2015, and Groom et al. 2013) but the
recommendations have not yet penetrated mainstream thinking in ornithological
circles. Some of these segmentation exercises have been carried out in eCognition
software (Nussbaum and Menz 2008), but this software is both specialised and
expensive and would not be readily available or accessible to most birdwatchers, in
direct contrast to ImageJ software.
Given suitable images to process, ImageJ can carry out relatively accurate automatic counts of large aggregations of birds. These automatic counts take seconds to
complete and come with three options for validation, (a) a numbered drawing of the
objects contributing to the count, (b) an Excel file listing the pixel sizes of each object
in the drawing and (c) partial or complete manual counts of the birds in the image.
C. Hurford