Search

This study quantifies and compares the accuracy of thematic habitat maps created from three types of remotely sensed data for Hawaiian near-shore coral reef environments. Remotely sensed data of similar spatial and temporal resolution were collected at four test areas chosen to represent a cross section of the coral reef habitat types present in the main Hawaiian Islands. Imagery derived from aerial photography, airborne AURORA hyperspectral data, and satellite IKONOS multispectral data was used to make maps by manual photointerpretation via "head's up" computer digitizing. Thematic map classes were based on a NOAA peer-reviewed hierarchical coral reef habitat classification scheme with four first-level major classes (unconsolidated sediments, submerged aquatic vegetation, coral reef and hard bottom, other delineations) and 28 detailed classes. For accuracy assessment, ground truth data were collected at 1225 stations selected by a stratified random sampling design. Each of the four test areas was sampled and tested independently, and the data were combined for a region-wide analysis. Accuracy assessment tests included error matrix analyses (overall, user's and producer's accuracy), the kappa statistic, the tau coefficient, and tests for significance differences between tau coefficients for maps generated from the three different types of imagery. The statistical accuracy assessment tests were preformed at the first level of the classification scheme. There were no significant differences between tau coefficients for maps generated from any of the three types of remotely sensed imagery at two of the four test areas (

The goal of this research is to develop a new approach to remote sensing thematic accuracy assessment in which the spectral distances between the classes in a thematic classification are used as inputs to the error estimation process. The conceptual basis for this new approach is that the confusion of relatively spectrally different classes represents a more severe error than confusing relatively spectrally similar classes. Therefore, the accuracy estimate of a classification can be adjusted to take into account the 'spectral severities' (or misclassification costs) of the errors in that classification. The benefits of including inter-class spectral distances in the accuracy assessment process are shown in the context of the development of two new accuracy assessment measures called Spectrally Weighted Kappa (SWK) and Spectrally Weighted Fuzzy (SWF). These two new accuracy assessment methods are introduced and tested for their performance relative to current techniques. The results of this research demonstrate that inter-class spectral distances can be used effectively in accuracy assessment of thematic classifications. The SWK approach can provide information about the spectral costs of errors in a classification that is not as apparent with traditional methods. In addition, SWK provides a quantitative base for establishing weights for Weighted Kappa analysis and allows for the possibility of improving a classification during its development. The SWF method improves upon current fuzzy accuracy assessment techniques by providing a way to establish membership functions that is based on inter-class spectral distances. We have shown that the SWF method can provide fuzzy membership values that are similar to those that a well-trained human might choose. Therefore, in cases where multiple interpreters would normally have been used to create fuzzy membership values, the SWF method can be employed reduce inter-interpreter bias. In addition, the SWF method provides a quantitative basis for establishment of fuzzy membership values. We expect that these two new accuracy estimation techniques will be of use to the remote sensing research community.