Genetic Programming with Convolutional Operators for Albatross Nest Detection from Satellite Imaging

Abstract

Conservation efforts in remote areas, such as population monitoring, are expensive and laborious. Recent advances in satellite resolution have made it possible to achieve sub-40 cm resolution and see small objects. This study aimed to count potential southern royal albatross nests based on the remote Campbell Island, 700 km south of New Zealand. The southern royal albatross population is declining, and due to its remoteness, there is an urgent need to develop new remote sensing methods for assessing the population. This paper proposes a new tree-based genetic programming (GP) approach for binary image segmentation by extracting shallow convolutional features. An ensemble of these GP segmentation models and individual GP models were compared with a state-of-the-art nnU-Net segmentation model trained on manually labelled images. The ensemble of shallow GP segmentation trees provided significantly more interpretable models, using <1% the number of convolutions while achieving performance similar to that of the nnU-Net models. Overall, the GP ensemble achieved a per-pixel F1 score of 75.44% and 123 out of 166 correctly identified nest-like points in the test set compared with the nnU-Net methods, which achieved a per-pixel F1 score of 74.49% and 129 out of 166 correctly identified nest-like points. This approach improves the practicality of machine learning and remote sensing for monitoring endangered species in hard-to-reach regions.