Construction of a geospatial predictor by fusion of global and local models

Geospatial data collected by remote sensing instruments are characterized by substantial variations in attribute values and relationships over space and time, posing great challenges to develop models with maximum predictive power. In this paper, we propose an approach in which global and local models are constructed, and predictions made by properly weighting their outputs. The algorithm is evaluated on aerosol optical thickness prediction using four consecutive MISR data sets collected in 2002 over the continental US. Results show that while the R² accuracy of the ANN global and local models are at most 0.25 and 0.4 respectively, the fusion model is significantly more successful, achieving R² accuracy above 0.50. In addition, accuracy improvements differ by spatial location, the largest being in the western US, and the smallest being in the east. This could be exploited to further improve the fusion algorithm.