Trial and Comparison of Some Encoder-Decoder Based Deep Learning Models for Automated Generalization of Buildings

  Objectives  The deep learning technology has made a breakthrough in many fields, and the automated map generalization based on the deep neural network is highly expected. Building generalization is a typical generalization case and can be abstracted and interpreted as the process of encoding and decoding. It is worth systematically studying the simulation of building generalization with encoder-decoder based deep neural networks.

  Methods  Samples are drawn from vector buildings via the segmentation of spatial data and the conversion between vector data and raster data, which is based on a natural principle of objective generalization (i.e., the legibility principle drawn by scale). The encoder-decoder based deep neural networks are trained and tested with these samples to learn buildings generalization and to evaluate this application. Five representative models based on encoder-decoder structure are constructed for learning of generali-zing buildings from 1∶10 000 to 1∶50 000. The experimental data cover 44.1 km×38.6 km and contain 89 539 and 21 732 buildings before and after generalization respectively.Five representative models are EDnet, Unet, ResUnet, Unet++ and Pix2Pix representing the basic multi-layers encoder-decoder neural network, the encoder-decoder neural network with skip connection, the residual encoder-decoder neural network, the dense encoder-decoder neural network, and the encoder-decoder neural network with generative adversarial mechanism respectively. The training process and test results of the three constructed models are compared, and effects of these five models while they are applied to the automated generalization of buildings are analyzed.

  Results  The pixel accuracies of predicted results during EDnet training and testing are lowest obviously, Unet takes the second lowest place, and the pixel accuracies of results generated by other models(ResUnet, Unet++ and Pix2Pix) are higher and relatively close. Additionally, there are some holes and fuzzy representations in generalized buildings predicted by ResUnet and Unet++, which does not conform with the regular representation of buildings. Generalization results predicted by the Pix2Pix show that most of generalized building edges are clear and generalized buildings can be represented independently and completely.

  Conclusions  The Pix2Pix is more suitable than the other four models for learning and simulating buildings generalization in our experiment. The encoder-decoder based deep neural networks are likely to deduce some knowledge and operators of buildings generalization by learning from samples of the original and generalized data. The encoder-decoder based deep neural networks have the potential in the learning of map generalization.