Automatically Designing U-Nets Using A Genetic Algorithm for Tree Image Segmentation

Tree image segmentation is an important task of analysing the forest cover in ecosystems. Convolutional neural networks (CNNs) particularly U-Nets are popular methods for image segmentation, but they often require rich expertise from both the problem domain and the neural network domain to design effective model architectures. Recently, automatically designing U-Nets has shown promise, but none of them has been proposed for tree image segmentation. When designing U-Nets, the encoding and decoding schemes are the keys to the success of the search, which have not been fully investigated yet. Therefore, this paper investigates a new approach based on genetic algorithms (GAs) to automatically design/learn U-N ets for tree segmentation from remote-sensing images. The new encoding scheme enables chromosomes of a GA to represent potential U-Net models of variable length based on blocks. The decoding scheme decodes chromosomes to potential U-Net models with the characteristics of U-Nets. More importantly, the new designs allow the new approach to effectively and efficiently search for U-Net models of variable lengths by only optimising several key parameters. The results show that the proposed approach achieves better performance than a well-known manually designed U-Net by automatically learning the U-Net model with a significantly smaller number of parameters on one tree image segmentation dataset.