Advanced Data Fusion Methods to Improve Wetland Classification Using Multi-Source Remotely Sensed Data

The goal of this research was to improve wetland classification accuracy and the reduction of classification errors and uncertainty by fully exploiting multi-source remotely sensed, and ancillary data and image metrics using advanced data analysis techniques. This PhD research executed in three phases: 1. Explorations of data type selections and significance in support of wetland classification. 2. The development of a hierarchically-based classification approach to best exploit the data identified and characterized through the first study. 3. The development of an ensemble classifier incorporating the aforementioned developments with Dempster-Shafer (D-S) theory in order to reduce errors and streamline computations.

The first phase explored the most effective data features, and metrics or families of data features in support of wetland classification. It was found that wetlands were best classified using the NDVI calculated from optical imagery obtained in the spring months, radar backscatter coefficients, surface temperature, and ancillary data such as surface slope, computed through either a Random Forest (RF) or Support Vector Machine (SVM) classifier. This work was also able to produce a wetland land cover map with an accuracy of 87.51% - an improvement from the ~82% typical of similar datasets and landcover types.

In the second phase a more effective approach to classify the aforementioned features in order to fully utilize the discriminant power of those features was explored. This was done through two hierarchically based classification strategies. The second hierarchically based RF classification methodology produced the most accurate classification result (91.94%).

The third phase focused on how to better exploit broad class separations and to reduce the propagation of errors and uncertainty which cascades through the classification hierarchy. These classifiers were integrated using D–S theory. Classification resulted in an overall accuracy of ~93% an improvement of 5% when compared to a traditional classification method. High level of confidence (>85%) misclassified pixels were reduced by ~10%.

The major contribution of this research was the improvement of classification accuracy and the reduction of classification errors and uncertainty through use of multiple classifiers, designed to best exploit broad class separations, through selected data features computed within a D-S framework.