Monitoring landscape changes by geographic information system and remote sensing.

Abstract

Landscape changes in Scio Township, Washtenaw County, southeastern Michigan from 1949 to 1985 were investigated using remote sensing and geographic information system (GIS) techniques, and predictive models for these changes constructed. Landscape patterns were delineated by photo interpretation and digitized. Landscape changes were detected with a map overlay method. Future land-use trends were predicted by four methods: multiple regression, Markov chains, multiple discriminant analysis (MDA), and multivariate nominal scale analysis (MNA). Most landscape changes resulted from displacement of agricultural lands by increasing urbanization, particularly development of residential, commercial, and industrial lands. Landscape changes tended to be spatially correlated, highly influenced by the highway construction and the development of Ann Arbor, where the University of Michigan resides. Spatial pattern analyses and indices of diversity and dominance suggest the landscape is becoming more fragmented. With a raster data format, optimal pixel size to represent Scio Township lies between 0.5 and 2.0 ha; and 1.0 ha was used. Temporal scale is also important; the longer the time interval, the greater the differences observed in the landscape pattern. Using multiple regression, loss of agricultural lands in areas close to I-94 are easier to predict than areas farther away. Most changes were at the expense of agricultural use. Some transition probabilities gave better forecasts of the landscape trends than the others using the Markov chain approach, although all correctly predicted change trends. One projection based on the transition probability matrix of 1949-1955 underestimated magnitudes of changes, while others overestimated. Prediction accuracy using MDA and MNA depends on the variables included. Using GIS-derived land-use information and environmental variables combined as predictors gave a better overall accuracy than either alone, but overall prediction accuracy never exceeded 74%. Environmental data can better predict agricultural and water land-uses, while GIS-derived land-use information is better at forecasting urban land, forests, and riverside. It was concluded that integration of GIS and remote sensing techniques is essential for investigating temporal and spatial patterns of landscape change. Several limitations in predicting future landscape patterns were found with all four statistical approaches examined.