Forecasting land-based environmental variables using similarity analysis and temporal graph convolutional neural networks

Abstract

Embodiments are directed to a computer-implemented method of analyzing a land region that has been decomposed into a plurality of regular or irregular sub-regions. The computer-implemented method includes applying, using a processor system, a feature extraction process that extracts a set of sub-region environmental descriptors for each of the plurality of sub-regions. The processor system applies a similarity analysis to the set of sub-region environmental descriptors to generate groups of the plurality of sub-regions. The processor system creates a plurality of group-based graphs by encoding each of the groups into a corresponding group-based graph. A spatio-temporal neural network is used to train a model based at least in part on the plurality of group-based graphs.