Computing and Information Systems - Theses
Permanent URI for this collection
Search Results
1 - 1 of 1
-
ItemEffective Spatial Feature Extraction via Convolutional Neural Networks( 2021)Spatial data analysis has achieved great success in many real-world applications such as region similarity learning, crime prediction, and traffic prediction. In spatial data analysis, spatial features play an important role. In this thesis, we propose novel techniques based on the Convolutional Neural Network (CNN) and its variants to learn spatial features from two important aspects, namely Point of Interest (POI) features and POI relationships. We focus on the building height feature, the hexagonal geometrical layout relationship, and the pairwise relationship of POIs. We show the importance of learning these spatial features in spatial data analysis applications, in particular, region similarity learning. We make the following contributions in this thesis: (1) We propose a CNN-based building height estimation method that computes building height from street scene images. We first detect roofline candidates from such images. Then, we use a deep neural network called RoofNet to filter these candidates and select the best candidate via an entropy-based ranking algorithm. When the true roofline is identified, we compute building height via the pinhole camera model. Experimental results show that our overall building height estimation method is more accurate than the baseline by up to 11.9%. (2) We propose a novel native hexagonal CNN framework named HexCNN for hexagonal layout relationship learning. HexCNN takes hexagon-shaped input and performs forward and backward propagation on the input based on hexagon-shaped filters, hence avoiding computation and memory overheads caused by transforming the input into rectangular shapes to fit traditional CNN models. Experimental results show that, compared with the state-of-the-art models, which imitate hexagonal processing but use rectangle-shaped filters, HexCNN reduces training time by up to 42.2%. Meanwhile, HexCNN saves memory by up to 25% and 41.7% for loading the input and performing convolution, respectively. (3) We propose a Graph Convolutional Network (GCN) model with weighted structural features named WGCN for learning edge direction features between POIs. WGCN first captures nodes' structural fingerprints via a direction and degree-aware Random Walk with Restart algorithm. The walk is guided by both edge direction and nodes' in- and out-degrees. Then, the interactions between nodes' structural fingerprints are used as the weighted node structural features. To further capture nodes' high-order dependencies and graph geometry, WGCN embeds graphs into a latent space to obtain nodes' latent neighbors and geometrical relationships. Based on nodes' geometrical relationships in the latent space, WGCN differentiates latent, in-neighbors, and out-neighbors with an attention-based geometrical aggregation. Experiments on transductive node classification tasks show that WGCN outperforms the baseline models consistently and by up to 17.07% in terms of accuracy on five benchmark datasets. (4) To showcase the effectiveness of learning the above spatial features in spatial data analysis applications, we propose a triplet-based Graph Neural Network model named C-MPGCN for region representation learning, where POIs are represented as graphs. C-MPGCN uses POIs' height and size as node features in addition to POIs' category and geo-location features used in existing methods. C-MPGCN also captures the POIs' hexagonal layout relationship and the pairwise relationship (edge features). Experimental results show that C-MPGCN outperforms the state-of-the-art methods for region similarity learning consistently under different evaluation matrices.