Polarimetric radar identification of invasive plant species in a prairie landscape

This thesis used synthetic aperture radar polarimetry to develop a methodology to identify invasive weed species in prairie rangeland using images acquired with Radarsat-2. It was part of a larger Government of Canada project to assess prairie rangeland. Twenty-four images at three incidence angles were collected of the Oldman river coulee north-west of Taber, Alberta, from the beginning of April to the end of October 2009. The images were processed using a Lee Sigma speckle filter followed by Freeman-Durden three component model-based decomposition and Cloude-Pottier eigenvalue-eigenvector decompositions. Training sets of different classes of vegetation were selected through the use of verified areas of vegetation. The Bhattacharyya distance and optimal Bayes error were used to determine theoretical class separability. A support vector machine scheme was used to separate the classes. The performance metric was assessed using the Cloude-Pottier, the Freeman-Durden, both decompositions together, and the full coherency matrix. The improvement in performance metric through temporal averaging, multiple incidence angles and temporal series was investigated. The theoretical separability was found to have a confidence of 90% for all non-brush classes while the brush classes had a confidence level of 75% using the coherency matrix, with the highest confidence level (80%) found using the steepest incidence angle. The temporal averaging gave an increase of 20% in performance metric over a single image for the Freeman-Durden and Cloude-Pottier combination parameter set and the coherency parameter set. The steeper incidence angles gave a consistently higher performance metric. When all three incidence angle images were combined the performance metric over the input parameter sets were very close to 100%. The temporal series gave an increase in performance metric over the single images, with the steepest incidence angle giving a performance metric of 100%. The performance metric increased as the number of input parameters increased. Among the different input parameter sets, the highest accuracies were seen for the combination of incidence angles. The performance metrics outside of the training sets could not be quantified and were limited to a visual inspection. Larger training sets would allow for independent training and validations sets.