Classification of Wetland Vegetation Based On TerraSAR-X Data: Comparison of Methods
Permanent address of the item is
Vesikasvillisuuden luokittelu TerraSAR-X sateliittidatan pohjalta: menetelmien vertailu
The aim of this Master’s thesis was to study the feasibility of different pattern recognition techniques for classification of wetland vegetation and ground cover from SAR images. Use of pattern recognition and remote sensing techniques provides an efficient and fast way to survey large areas of earth’s surface without time consuming and difficult field studies. Utilization of SAR-technology enables monitoring of the area regardless of time or weather conditions unlike passive radiometers operating in visible light spectrum. The thesis is part of the biosphere modelling project of Posiva Oy, the instance responsible for spent nuclear fuel disposal of the nuclear power plant in Olkiluoto. The test data for the classification algorithms consisted of a SAR image obtained from Lake Poosjärvi on 4th of July 2014. Lake Poosjärvi is a small (area of 794 hectares) lake located in the western part of Finland in the Satakunta region. Forest type at Lake Poosjärvi is mainly mixed forest, and ground vegetation is dominated by sedges, water horsetail, water lilies, pondweeds and reed. Seven classes were used in classification: reed, reed in water, water horsetail, water, trees, sedge and yellow water lily. In the thesis, the feasibility of four different classification algorithms was studied: Random Forest, Support Vector Machine, Multilayer Perceptron ensemble and K-Nearest Neighbor. Features used in classification were selected from image processing and remote sensing literature. It was found, that the classification methods used in the thesis were applicable to the task, and that texture-based features had significant positive im-pact on classification results. The best classification results were achieved using the Multilayer Perceptron ensemble classifier, though Random Forest and Support Vector Machine produced good results as well. The question whether the classifiers were classifying the test data according to the properties of soil, ground cover or vegetation, remains open.