Automated Detection of Male Eiders over Multispectral Aerial Photographs

An automated method has been developed to detect male eiders (COEI_M) on aerial multispectral images. Potential bird regions were determined after filtering the images with an average filter, masking out the ocean background and the land areas, and applying a watershed segmentation algorithm. Large and small objects were then removed. For each remaining object and its neighborhood, spectral and geometrical variables were computed. The variables that were not significantly different between the true and false COEI_M classes were removed based on a p-value of higher than 0.08 computed by an ANOVA. The objects were divided into training and validation datasets. A support vector machine using a radial basis function was used to classify the objects in the two classes. We achieved an overall accuracy of 98.6% with the validation data set. The user's accuracy and producer's accuracy for the true COEI_M class in the validation dataset were 86.4% and 92.7%, respectively.