Land Cover Change Modeling Using Cellular Automata Rules Derived from Landsat Imagery

Land cover generally describes categorical feature classes defined by their physical characteristics, such as vegetation or material type, as seen on a small parcel of surface area of the Earth. Land cover changes occur annually due to such activities as urban development, agriculture, climatic shifts, and natural disasters such as fires and hurricanes. Observing land cover change provides insight into trends due to natural and manmade annual changes that can be used to construct decision rules and to model techniques purposed with preventing or mitigating the effects of extreme weather or climatic shifts.

Annual datasets are required for such analysis and the United States Geological Survey (USGS) provides land cover datasets covering the entire United States for only the years 1992, 2001, 2006, and 2011, which cover a time frame of 19 years. One mapping technique called Variable Multiple Endmember Mixture Analysis (VMESMA) accurately maps physical characteristics of land but can be computationally intensive and slow depending on its implementation. In this thesis, Variable Spectral Unmixing (VSU) — a new and improved spectral mixture analysis technique inspired by VMESMA — is presented to produce land cover estimates from preprocessed Landsat imagery for the years 2001 through 2011. VSU results correspond to physical surface material types, such as coniferous trees and artificial substances, and are interpreted into land cover classes based on material type prior to overall classification by hierarchal rules. Agreements with the USGS National Land Cover Dataset (NLCD) of less than 40% result due to the classification rules and reflect the physical surface types that meet the first rule within the hierarchy. Future land cover mapping applications require new classification rules to improve interpretation of the VSU results and the agreement of the generated maps with the USGS NLCD.

Land cover estimates are used to develop Cellular Atomaton-based (CA) decision rules to map land cover change and to forecast such changes into future years. The CA rules are based on the analysis of Moore and von Neumann neighborhoods of a time series of VSU-generated maps. Results of the neighborhood analysis revealed potential general neighborhood structures for decision rules, which may or may not vary in time as a result to changes in the rates of change of each class. Forecast results are tested in a basic iterative fashion using the USGS NLCD 2001 map as a base case. Agreements of 66% and 77% of the von Neumann and Moore forecasts, respectively, for the year 2011 with the NLCD 2011 demonstrate the feasibility of land cover change modeling using neighborhood-based CA decision rules and a method for modeling land cover change trends based on decision rules derived from a time series of maps.