This thesis aimed to combine experts’ domain knowledge and deep learning models to solve practical problems in different civil and environmental engineering applications. The first application is bird watching for environmental impact assessment, and the second one is drainage network extraction. For bird watching, the domain randomization strategy was adopted to enhance the accuracy of the deep learning models in bird detection and classification. Trained with virtual bird images with sufficient variations in different environments, which were generated from 3D bird models created based on ornithological domain knowledge, the model was guided to focus on the fine-grained features of birds and achieve higher accuracies. In addition, semi-supervised learning, which used both the high-qua...[ Read more ]

This thesis aimed to combine experts’ domain knowledge and deep learning models to solve practical problems in different civil and environmental engineering applications. The first application is bird watching for environmental impact assessment, and the second one is drainage network extraction. For bird watching, the domain randomization strategy was adopted to enhance the accuracy of the deep learning models in bird detection and classification. Trained with virtual bird images with sufficient variations in different environments, which were generated from 3D bird models created based on ornithological domain knowledge, the model was guided to focus on the fine-grained features of birds and achieve higher accuracies. In addition, semi-supervised learning, which used both the high-quality bird images labeled by experts and low-quality unlabeled bird images collected at the study site, was also explored to improve the efficiency and efficacy of bird watching with the aid of deep learning techniques. Specifically, the (K + 1)-class discriminator-based Generative Adversarial Network (GAN), a semi-supervised model, was proven effective for disentangling the species and species-independent information into two orthogonal linear spaces. Based on this result, the GAN enhanced with the Orthogonal Weights (OW-GAN) approach was proposed to extract the species-independent information and classify different bird species without the influence from the species-independent perturbations. Using these techniques developed, continuous observations on the activities of egrets at Penfold Park, Hong Kong, where the bird images were captured using a tailor-made 360-degree camera, were conducted for an extended period at high sampling frequencies, without the survey constraints when done by human experts. The analyzed results were not only consistent with manual observations in terms of the accuracy in bird detections and classifications, but also revealed new insights on egret behavior based on long-term monitoring. For drainage network extraction, distributed representations of aspect features were extracted, with the help of the Deterministic 8 (D8) algorithm that was proposed by human experts, to facilitate the deep learning model for calculating the flow direction. A semantic segmentation model, U-Net, was then adopted to improve the accuracy and efficiency in both predicting the flow directions and in pixel classifications (i.e., waterbody, flowline, or hillslope). Subsequently, postprocessing was used to enhance the flowline delineation. The proposed framework achieved state-of-the-art results, including the required computation time, compared with the traditional methods and the published deep-learning-based methods. Further, case study results demonstrated that the proposed framework could extract drainage networks with high accuracy for rivers of different widths flowing through different types of terrains, even with roads and dams. This framework, requiring no parameters to be provided by users, can also produce waterbody polygons and allow cyclic graphs in the drainage network.