Raster based coastal marsh classification within the Galveston Bay ecosystem, Texas

A mapping study using remote sensing software called ENVI was conducted utilizing four software algorithms to investigate whether these techniques could accurately classify habitat types and vegetation communities along West Bay of the Galveston Bay Ecosystem from color infra-red (CIR) imagery. The algorithms were used in a small-scale study to investigate which of these techniques could most accurately distinguish habitat types and vegetation communities from the imagery at a site specific location. The most accurate algorithm of the four was used in a large-scale classification study in which entire images were classified utilizing the same data from the small-scale study. Regions of interest (ROIs) were used within ENVI to specify areas of interest within each image that was classified. The locations of ROIs were recorded using a GPS prior to classification, then each was added into ENVI as data points, and each ROI polygon was digitized according to its respective pixel color. Once all of the ROI polygons were completed, each software algorithm was employed. After classification, each habitat type and vegetation community was ground-truthed in order to verify the accuracy of the algorithms. The position points were added as ground truth points within ENVI and an accuracy matrix was assessed. The technique with the greatest averaged accuracy within the smallscale study was selected for the large-scale study. The ROIs and ground truth points used in the smallscale study were used again in the large-scale study. The small-scale study concluded that the Parallelepiped algorithm produced significantly less accurate classifications than the other three. Although the Mahalanobis algorithm was not significantly different from the other two algorithms, it yielded the highest overall average accuracy and was used in the large-scale study. In both the small-scale and large-scale studies there was no significant difference in the two different years of aerial imagery and there were no significant differences in accuracy for locations. None of the software algorithms were accurate at classifying habitat types and vegetation communities using the imagery. The accuracy for the Mahalanobis algorithm was less than 60%. Inaccuracies were largely due to overlapping spectral signatures among habitat types and vegetation communities.