Haven C. Sweet


Conventional image analyzing techniques are inaccurate when applied to Landsat data obtained from highly heterogeneous regions. In much of Florida, plant communities occupy small areas with erratic boundaries, meaning that most Landsat pixels represent a mixture of several different plant formations. This is especially true when studying an alien tree such as Melaleuca which is expanding its range by infiltrating a variety of plant communities. This paper describes development of a technique for using a personal computer to identify and quantify the major communities (plus sand and water) within a single pixel, and applies it to attempts to inventory the spread of Melaleuca.

High resolution photographs were used to prepare a vegetation map of all pure plant communities and dense Melaleuca stands in a 14km2 study site. Landsat data was manipulated with the Image 100 until the spectral limits of dense Melaleuca stands, Flatwoods, Cypress, Swale, Sand and Water produced a coverage map that agreed with the ground truth map. Using the entire Landsat scene, each spectral band of all pixels identified as being one of the pure community forms was averaged to produce six sets of four point spectra (library spectra).

Pixel data from the study site was hand entered on an Apple II computer where a Pascal program determined what proportions of the six library spectra provided the best least squares fit of the unknown pixel's spectrum. Since some spectra could yield several different solutions, the program forced a number of solutions, each accompanied by three measurements of the solution's error. The program evaluated the error terms of all solutions, printing that solution with the least error; if all solutions exceeded acceptable limits, the pixel was labeled unsolvable.

Comparison of the computer's assessment of sixty pixels with that observed in aerial photos of an equivalent area was encouraging. Melaleuca was found to have an average coverage of 54.7% with the Apple as opposed to 55.6% as estimated from aerial photos. In addition, communities which were minor components of many pixels were fairly accurately measured (Sand, 13.2% by Apple vs. 14.9% by photo; Flatwood, 11.4% vs 16.8%; Swale, 7.1% vs 5.7%; Water, 4.4% vs. 3.2%; and Cypress, 1.6% vs 4.2%). Three of sixty pixels were unsolvable.

Date of this Version