Deep Learning-Based Spatial Information Prediction

Authors: Wonho Jo*,
Topics: Geographic Information Science and Systems
Keywords: Convolutional Neural Network, LSTM, Time series, Spatial data
Session Type: Paper
Presentation File: No File Uploaded


Traditional prediction problems rely on time series forecasting models. For time series prediction, autoregressive integrated moving average(ARIMA) and its variants have been mainly used. However, it cannot model the spatio-temporal structures. Deep learning-based information prediction has two big category. One is Convolutional Neural Network which can find spatial relation. This Convolutional Neural Network method is motivated by the First Law of Geography. The other one is Recurrent Neural Network, especially LSTM(Long-short-term-memory) which can consider temporal relation. This study proposes the Convolutional-LSTM which is combination of two models for spatial information prediction. This study uses time series raster data for the prediction. Given a set of historical raster images, this model can predict the spatial information for next time step. Specific time interval have to be set for prediction. Convolutional-LSTM find the spatial relationship for each time lag and analyzes them in a time-wise fashion for temporal relation. Consequently, Convolutional-LSTM captures the complex nonlinear relations of both space and time. This paper focuses on PM10 prediction as a case study to investigate the feasibility of the proposed model for spatio-temporal information prediction.

Abstract Information

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

To access contact information login