Date of Award


Document Type


Degree Name

Doctor of Philosophy (PhD)

First Advisor

Nina N. S.-N. Lam


The rapid decline in acreage of land areas in wetlands caused by frequent inundations and flooding has brought about an increased awareness and emphasis on the identification and inventory of land and water areas. This dissertation evaluates three classification methods--Normalized Difference Vegetation Index technique, Artificial Neural Networks, and Maximum-Likelihood classifier for the delineation of land/water interface conditions using Landsat-TM imagery. The effects of three scaling algorithms, including resampling by aggregation, Gaussian smoothing, and local variance analysis, on the classification accuracy are analyzed to determine how the delineation, quantification and analysis of land/water boundaries relate to problems of mixed pixels, scale and resolution. Bands 3, 4, and 5 of a Landsat TM image from Huntsville, Alabama were used as a multispectral data set, and ancillary data included USGS 7.5 minute Digital Line Graphs for classification accuracy assessment. The 30 m resolution multispectral imagery was used as baseline data and the images were degraded to a series of resolution levels and Gaussian smoothed through various scaling constants to simulate images of coarser resolution. Local variance was applied at each aggregation and scaling level to analyze the textural pattern. Classifications were then performed to delineate land/water interface conditions. To study effects of scale and resolution on the land/water boundaries delineated, overall percent classification accuracies, fractal analysis (area-perimeter relationships), and lacunarity analysis were applied to identify the range of spatial resolutions within which land/water boundaries were scale dependent. Results from maximum-likelihood classifier indicate that the method marginally produced higher overall accuracies than either NDVI or neural network methods. Effects from applying the three scaling algorithms indicate that overall classification accuracies decrease with coarser resolution, increase marginally with scaling constant, and vary non-linearly with local variance mask sizes. It was discovered that the application of Gaussian smoothing to neural network classifier produces very encouraging results in classifying the transition zone between land and water (mixed pixels) areas. Fractal analysis on the classified images indicates that coarser resolutions, higher scaling constants and higher degrees of complexity, wiggliness or contortion of the perimeter of water polygons span higher ranges of fractal dimension. As the water polygons become more complex, the perimeter becomes increasingly plane filling. From the changes in fractal dimension, lacunarity analysis and local variance analysis, it is observed that at 150 m, a peak value of measured index is obtained, before dropping off. This suggests that at 150 m, the aggregated water bodies shift to a different 'characteristic' scale and the water features formed are smooth, compact, have more regular boundaries and form connected regions. This scale dependence phenomenon can help to optimize efficient data resampling methodologies.