- Liou, Wen-Shu
University of Maine
The objectives of this research were to analyze, revise, and test a Geographic Information System (GIS) rule-based model to improve the mapping of forested wetlands in Maine. To determine the performance of satellite mapping techniques, three conventional image classification methods (unsupervised, tasseled cap, and hybrid classification) and a revised GIS rule-based model were evaluated. Accuracy assessment was conducted by cross tabulation of classification results to photo interpreted random sample plots. The accuracy of conventional methods revealed very similar results between study sites. Overall accuracy for four super classes (forested wetland, other wetland, forested upland, and other upland) ranged from 74% to 78% at the Acadia study site and 72% to 81% at the Orono study site.
The GIS model incorporated hydric soils, slope, National Wetland Inventory, and hydrography data. In order to clarify the contribution of GIS variables to forested wetland delineation, several analytical methods were conducted. After the GIS layers were analyzed statistically, a integrated model was formulated. The new integrated GIS model offered a greater degree of versatility and automation for a less subjective classification approach in the forested wetland mapping application.
The integrated GIS model was tested using the Orono data set. The results indicated that the model had the highest classification accuracy among all tested methods at both study sites. The overall accuracy and Kappa correction for the four classes were 82% and 0.70 at the Acadia study site and 81% and 0.72 at the Orono study site. Pairwise significance tests indicated that the integrated model was significantly better than unsupervised and tasseled cap classification methods at both study sites. The Kappa coefficients for the integrated model were slightly higher compared to the hybrid classification approach, however the significance test indicated no difference between the two methods at either study site.
Hydric soils, slope, and National Wetland Inventory data were the most important variables in the integrated GIS model. However, the relative importance of the model variables in predicting wetland conditions differed between study sites. The results suggest that the physical characteristics of the two study sites may have had more influence on the conventional classification methods than on the integrated model because the model incorporated the physical variables into the decision rule. This investigation agrees with Ahl (1994) in that supervised or hybrid classification techniques and a GIS model are the two most promising techniques for further investigations of forested wetland mapping in Maine.
Evaluation of new variables in the model (e.g. topographic position) and the effect of spatial error propagation in developing the GIS data base need to be investigated further.