Spatial interpolation based on conditional generative adversarial neural network

Authors: Di Zhu*, Peking University, Ximeng Cheng, Peking University, Fan Zhang, Massachusetts Institute of Technology, Yong Gao, Peking University, Yu Liu, Peking University
Topics: Applied Geography, Geographic Information Science and Systems, Spatial Analysis & Modeling
Keywords: spatial interpolation; generative adversarial networks; deep learning; encoder-decoder; spatial sampling
Session Type: Paper
Day: 4/5/2019
Start / End Time: 5:00 PM / 6:40 PM
Room: Washington 6, Marriott, Exhibition Level
Presentation File: No File Uploaded


Spatial interpolation is a traditional geostatistical operation that aims at predicting the attributes of unobserved locations given a sample of data defined on point supports. Deep neural networks, especially the idea of generative adversarial networks (GANs), provides us a novel perspective to implement spatial interpolation with their outstanding performance in learning representation from high dimensional data. In this research, we designed a conditional encoder-decoder generative adversarial network (CEDGAN) that can learn the deep feature maps of spatial dependency and infer the values at unobserved locations based on observed ones. Utilizing a dataset of digital elevation models (DEM) that contains various terrains in China as example, we trained the proposed network through an adversarial game between the generator and the discriminator. The interpolation error can reach near 2.5 meters per location when the spatial sampling ratio is lower than 10\%. The proposed method excels benchmark methods in aspects of accuracy, batching capability, computing speed and visual fidelity. Our work demonstrates the feasibility of spatial interpolation based on conditional generative neural networks and enlightens a new branch of traditional spatial interpolation methods in various geographic applications.

Abstract Information

This abstract is already part of a session. View the session here.

To access contact information login