This dissertation explores the intersection of Geographic Information Science (GIScience) and Artificial Intelligence (AI), specifically focusing on the enhancement of 3D deep learning models by spatial principles for understanding 3D geospatial data. With the rapid advancement in geospatial technologies and the proliferation of 3D data acquisition methods, there is a growing necessity to improve the capability of AI models to interpret complex 3D geospatial data effectively. This work seeks to leverage spatial principles, particularly spatial autocorrelation, to address the challenges pertaining to 3D geospatial object detection.
The research is structured around three pivotal questions: the utility of spatial autocorrelation features for understanding 3D geospatial data, the approach to derive content-adaptive spatial autocorrelation features, and the enhancement of post-processing in the task of 3D geospatial object detection. Through a series of experiments and model developments, this dissertation demonstrates that incorporating spatial autocorrelation features, such as semivariance, significantly enhances the performance of 3D deep learning models in geospatial object detection. A novel spatial autocorrelation encoder is introduced, integrating spatial contextual features into the 3D deep learning workflow and thereby improving accuracy in detecting objects within complex urban and natural environments. Further, the dissertation delves into the challenges brought by data partitioning and sampling in large-scale 3D point clouds, as evidenced in the DeepHyd project focusing on the detection of hydraulic structures (i.e., bridge and its components). The findings highlight the critical role of spatial dependency patterns in optimizing object detection accuracy and pave the way for future improvement of the 3D deep learning frameworks.