Spatio-temporal forecasting using wavelet transform-based decision trees with application to air quality and covid-19...

Area of Science:

• Data Science
• Environmental Science
• Epidemiology

Background:

• Time series forecasting often struggles with spatial spillover effects.
• Decision trees offer interpretability but can be limited in handling complex temporal and spatial dependencies.
• Wavelet transforms can represent data at various resolutions, potentially enhancing feature extraction.

Purpose of the Study:

• To develop a hybrid decision tree and wavelet transform method for time series forecasting.
• To improve prediction accuracy and interpretability of time series with spatial spillover effects.
• To apply and validate the method on simulated, air pollution, and COVID-19 data.

Main Methods:

• A decision tree model integrated with wavelet transform for feature extraction.
• Application of Haar, LA8, D4, and D6 wavelets for time series decomposition.
• Construction of spatial weights to model spillover effects in contiguous regions.

Main Results:

• The Haar wavelet demonstrated superior performance in simulations.
• The hybrid model successfully identified autoregressive, seasonal, and spatial spillover effects in air quality index data.
• Wavelet-transformed variables yielded improved forecasting performance and interpretability compared to original data.

Conclusions:

• The developed method effectively forecasts time series data with spatial spillover effects.
• Wavelet transform enhances decision tree performance and interpretability for complex time series.
• Analysis of COVID-19 data suggests lockdown policies were effective, as indicated by the non-selection of spatial weighted variables.