Data‐driven polyline simplification using a stacked autoencoder‐based deep neural network

Abstract

Automatic simplification of polylines is an important issue in spatial database and mapping. Traditional rule‐based methods are usually limited in performance, especially when the man‐made rules have to be adapted to different polylines with different shapes and structures. Compared to the existing neural network methods focusing only on the output layer or the code layers for classification or regression, our proposed method generates multi‐level abstractions of polylines by extracting features from multiple hidden layers. Specifically, we first organize the cartographic polylines into the form of feature vectors acceptable to the neural network model. Then, a stacked autoencoder‐based deep neural network model is trained to learn the pattern features of polyline bends and omit unimportant details layer by layer. Finally, the multi‐level abstractions of input polylines are generated from different hidden layers of a single model. The experimental results demonstrate that, compared with the classic Douglas–Peucker and Wang and Muller algorithms, the proposed method is able to properly simplify the polylines while representing their essential shapes smoothly and reducing areal displacement.

Integrating domain knowledge and graph convolutional neural networks to support river network selection

“…Specifically, instead of using river network knowledge, the data‐driven GCNN utilized the embedding of reach coordinates as the node attributes of the DG_RN. The reach coordinates were fed into an Autoencoder for unsupervised learning, and the encoder's output was used as the reach coordinate embedding (Yu & Chen, 2022). These embeddings replaced the knowledge (i.e., Strahler , NumT , DPR , and UDA ) as the node features of the DG_RN.…”

“…For instance, building generalization was implemented using a deep convolutional neural network (Feng et al, 2019), while polyline generalization (Du et al, 2022), such as mountain road generalization (Courtial et al, 2022) and topographic map generalization in urban areas (Courtial et al, 2021), was performed using a generative adversarial network. Nevertheless, methods such as end-to-end multiple autoencoders for polyline generalization (Yu & Chen, 2022) or graph convolutional networks for road network selection (Zheng et al, 2021) have certain limitations. For example, fixed-length data input and transductive learning can only be applied to specific samples, which reduces their robustness and applicability in different scenarios.…”

“…The first step is to reconstruct vector data (e.g., points, polylines, or polygons) commonly used in map generalization to align with the deep learning data paradigm. Current reconstruction methods for specific problems include using a triangulation network structure for point cloud data classification and semantic segmentation (Bazazian & Nahata, 2020; Jiang et al, 2021), a coordinate sequence structure for line simplification (Yu & Chen, 2022), and a dual graph of drainage and a minimum binary tree (Yu, Ai, Yang, Huang, & Gao, 2023; Yu et al, 2022) of building for pattern recognition (Yan et al, 2019). However, map generalization involves multiple operators, such as selection, simplification, and aggregation, and different objects like bends or nodes of rivers, roads, and boundary lines for simplification; arcs or strokes of river and road networks for selection; and polygons of buildings and lakes for aggregation.…”