Master's lecture in Software Engineering - Erwin Szudrawski
![](https://english.hi.is/sites/default/files/styles/efsta_mynd___vi_bur_um/public/stjori/bakgrunnsmyndir/googlemap_3.png?itok=vuT5ym9A)
Tæknigarður
Room 227
Master's student: Erwin Szudrawski
Title: Generation of Training Data for Automatic Land Cover Classification with Machine Learning
___________________________________________
Faculty: Faculty of Industrial Engineering, Mechanical Engineering and Computer Science
Advisors: Helmut Neukirchen,Professor at the Faculty of Industrial Engineering, Mechanical Engineering and Computer Science and Ernir Erlingsson
Examiner: Gabriele Cavallaro, Post doc at Forschungszentrum Jülich, Germany
Abstract
The importance of monitoring change in land cover has been increasing with time and technological development. Urbanization, agriculture, deforestation and livestock grazing are primary factors influencing this change which can have devastating consequences if not monitored properly. Thematic datasets, e.g. Corine Land Cover (CLC) for Europe, are usually updated manually, through visual interpretation of remote sensing images, which is inefficient from the perspective of cost and time. The manual process can be automated with machine learning algorithms,
such as Fully Convolutional Neural Network (FCN), but they require substantial amount of data in order to produce satisfactory results. Rare land cover classes are especially problematic due to class imbalance (which is almost always present in the environmental data) and coarse resolutions of the available thematic datasets. The goal of this thesis is to develop software that simplifies the choice of the training data in order to improve classification results on rare classes, as well as automating the refinement and preparation process of the data. The first goal is achieved by developing a visual tool that enables the user to easily see the land cover ratio in the region of their choice. The tool was evaluated on a region in Poland and increased the F-score of wetlands from 0% to 70%. The other goal is achieved by a simple program developed in Python, that performs the refinement and generation process automatically. The use of this software could have a positive impact on the quality and efficiency of future research in this area.