285
Random Forests. A random forest is a collection of classification trees trained on a
subsample of the training data (Chen 2007). Random Forests are currently one of the
top performing algorithms for data classification and regression. They are widely used
because of their ability to classify large amounts of data with high accuracy.
In both cases, we used chicks and adult gulls from the orthomosaic as training
areas for the class “bird” by selecting body pixels as in Grenzdörffer (2013). The
rest of the training classes represented open water, vegetation and bare soil. Finally,
where the number of pixels associated with a bird was less than a certain threshold,
these were removed from the classification. We carried out an accuracy assessment
against the photo-interpretation results for the whole orthomosaic. Swimming birds
were excluded.
Colony Monitoring
Ground-based monitoring surveys comprised periodic visits to the colony during
the breeding season and the estimation of pairs using binoculars and telescopes
from the closest position to the island to avoid disturbance. In a separate one-off
visit during the breeding season, the monitoring team entered the colony to estimate
the number of nests and to ring the chicks brought offshore. The date on which the
colony was accessed was determined and informed by the continuous observations
of the breeding birds.
Results
Visual Identification of Birds
Visual identification of the birds took about three days of work for each orthomosaic
(Fig. 5). An experienced ornithologist used one working day to identify birds and a
second day to label them according to their activity (flying, lying, standing up, etc.).
Finally, on the third day, the supervisor assessed the identification and labelled the
individuals. As C. genei nests are not visible, we assigned nests to lying birds.
Table 2 shows the results for UAV flight #2. That was the flight-date when most of
the colony was lying: we selected this flight as the reference for 2015. A total of 915
individuals of Slender-billed Gull were identified together with 149 individuals of
Black-headed Gull. Another 27 birds were geolocated in the orthomosaic, corresponding to several individuals of White-headed Duck (Oxyura leucocephala)
and Yellow-legged Gull (Larus michahellis). In addition, 97 visible eggs from 47
nests (up to 4 eggs per nest) were identified and labelled. UAV flight #3 took place
late in the season and, by this time, many chicks were already grouped in nurseries.
Therefore, for this flight and in addition to adults, we also identified chicks that were
either alive or dead (Table 3).
Using UAVs to Map Aquatic Bird Colonies
Random Forests. A random forest is a collection of classification trees trained on a
subsample of the training data (Chen 2007). Random Forests are currently one of the
top performing algorithms for data classification and regression. They are widely used
because of their ability to classify large amounts of data with high accuracy.
In both cases, we used chicks and adult gulls from the orthomosaic as training
areas for the class “bird” by selecting body pixels as in Grenzdörffer (2013). The
rest of the training classes represented open water, vegetation and bare soil. Finally,
where the number of pixels associated with a bird was less than a certain threshold,
these were removed from the classification. We carried out an accuracy assessment
against the photo-interpretation results for the whole orthomosaic. Swimming birds
were excluded.
Colony Monitoring
Ground-based monitoring surveys comprised periodic visits to the colony during
the breeding season and the estimation of pairs using binoculars and telescopes
from the closest position to the island to avoid disturbance. In a separate one-off
visit during the breeding season, the monitoring team entered the colony to estimate
the number of nests and to ring the chicks brought offshore. The date on which the
colony was accessed was determined and informed by the continuous observations
of the breeding birds.
Results
Visual Identification of Birds
Visual identification of the birds took about three days of work for each orthomosaic
(Fig. 5). An experienced ornithologist used one working day to identify birds and a
second day to label them according to their activity (flying, lying, standing up, etc.).
Finally, on the third day, the supervisor assessed the identification and labelled the
individuals. As C. genei nests are not visible, we assigned nests to lying birds.
Table 2 shows the results for UAV flight #2. That was the flight-date when most of
the colony was lying: we selected this flight as the reference for 2015. A total of 915
individuals of Slender-billed Gull were identified together with 149 individuals of
Black-headed Gull. Another 27 birds were geolocated in the orthomosaic, corresponding to several individuals of White-headed Duck (Oxyura leucocephala)
and Yellow-legged Gull (Larus michahellis). In addition, 97 visible eggs from 47
nests (up to 4 eggs per nest) were identified and labelled. UAV flight #3 took place
late in the season and, by this time, many chicks were already grouped in nurseries.
Therefore, for this flight and in addition to adults, we also identified chicks that were
either alive or dead (Table 3).
Using UAVs to Map Aquatic Bird Colonies