Predicting spatial data with RBF networks

Area of Science:

• Geospatial analysis
• Machine learning

Background:

• Conventional radial basis function (RBF) networks assume independent and identically distributed data, limiting their effectiveness in spatial prediction tasks.
• Spatial prediction inherently requires incorporating spatial information, which standard RBFs do not adequately address.

Purpose of the Study:

• To improve spatial prediction accuracy by fusing spatial information within RBF networks.
• To investigate the optimal layer for spatial information fusion in RBF architectures.

Main Methods:

• Modified RBF networks by incorporating spatial information at various network layers.
• Conducted experiments to evaluate the performance of RBF networks with different spatial information fusion strategies.
• Analyzed the impact of a coefficient used in the output layer's linear combination.

Main Results:

• Spatial information fusion significantly improves RBF network performance for spatial prediction.
• Fusion of spatial information at the hidden layer demonstrated the most effective results.
• An optimal coefficient value of approximately one was identified for the output layer combination.

Conclusions:

• Integrating spatial information into RBF networks is crucial for accurate spatial prediction.
• The hidden layer is the optimal location for spatial information fusion within RBF architectures.
• Further research can explore the precise role and optimization of the output layer coefficient.