Information-Theoretic Modeling of Categorical Spatiotemporal GIS Data

Area of Science:

• Geographic Information Systems (GIS)
• Data Mining
• Information Theory

Background:

• Categorical spatiotemporal land use data present analysis challenges.
• The National Land Cover Database (NLCD) provides discrete, gridded land cover data from 2001 to 2021.
• Evergreen Forest (EFO) is identified as the most dynamic land cover class in the study area.

Purpose of the Study:

• To apply an information-theoretic data mining method, Reconstructability Analysis (RA), to model and predict categorical spatiotemporal GIS land use data.
• To determine the predictive accuracy of RA for the presence or absence of Evergreen Forest (EFO) in 2021.
• To interpret the model's findings in the context of land use dynamics and forest clear-cutting cycles.

Main Methods:

• Utilized Reconstructability Analysis (RA), a maximum-entropy-based methodology for discrete data.
• Employed NLCD data from 2001-2021, focusing on EFO as the dependent variable.
• Modeled predictions using a sparse set of cells from a spacetime data cube of neighboring lagged-time cells.

Main Results:

• RA achieved approximately 80% accuracy in predicting the 2021 EFO land cover state.
• The presence of Shrubs and Grasses in preceding time steps indicated a high probability of 2021 EFO.
• The presence of EFO in preceding time steps indicated a high probability of its absence in 2021.

Conclusions:

• The findings suggest RA can detect forest clear-cut cycles, explaining the dynamism of the EFO class.
• This study introduces a novel approach for analyzing categorical GIS data using an entropy-based methodology.
• The RA methodology demonstrates successful application in modeling complex spatiotemporal geographic information systems data.