Encoder-Decoder Full Residual Deep Networks for Robust Regression and Spatiotemporal Estimation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning
• Environmental Science

Background:

• Deep neural networks can model complex nonlinear relationships but suffer from accuracy degradation due to vanishing gradients.
• This limits their application in predicting continuous variables with small sample sizes or weak translation invariance.
• Existing methods struggle with spatiotemporal imputation and estimation tasks involving significant missing data.

Purpose of the Study:

• To develop a robust deep learning network that overcomes accuracy degradation in complex spatiotemporal variable prediction.
• To enhance signal propagation and learning outcomes using full residual connections.
• To provide a powerful tool for applications involving nonlinear relationships, varying sample sizes, and spatiotemporal data.

Main Methods:

• Developed an encoder-decoder full residual deep network inspired by computer vision architectures.
• Embedded full shortcuts from each encoding layer to its corresponding decoding layer.
• Systematic encoder-decoder architecture facilitates efficient residual mapping and error signal propagation.

Main Results:

• The proposed network structure with full residual connections boosts backpropagation and improves learning outcomes.
• Achieved state-of-the-art accuracy in spatiotemporal imputation of aerosol optical depth and estimation of PM2.5.
• Demonstrated less bias in predicting extreme values and generated more realistic spatial surfaces compared to traditional methods.

Conclusions:

• The encoder-decoder full residual deep network is an effective solution for complex spatiotemporal predictions.
• The method shows superior performance in handling data with missingness and weak translation invariance.
• This novel approach offers a powerful tool for diverse applications requiring accurate continuous variable prediction.