Automatic Detection of Low Passability Terrain Features in the Scandinavian Mountains

Sammanfattning: During recent years, much focus have been put on replacing time consuming manual mappingand classification tasks with automatic methods, having minimal human interaction. Now it ispossible to quickly classify land cover and terrain features covering large areas to a digital formatand with a high accuracy. This can be achieved using nothing but remote sensing techniques,which provide a far more sustainable process and product. Still, some terrain features do not havean established methodology for high quality automatic mapping.The Scandinavian Mountains contain several terrain features with low passability, such asmires, shrub and stony ground. It would be of interest to anyone passing the land to avoid theseareas. However, they are not sufficiently mapped in current map products.The aim of this thesis was to find a methodology to classify and map these terrain featuresin the Scandinavian Mountains with high accuracy and minimal human interaction, using remotesensing techniques. The study area chosen for the analysis is a large valley and mountain sidesouth-east of the small town Abisko in northern Sweden, which contain clearly visible samplesof the targeted terrain features. The methodology was based on training a Fuzzy Logic classifierusing labeled training samples and descriptors derived from ortophotos, LiDAR data and currentmap products, chosen to separate the classes from each other by their characteristics. Firstly,a set of candidate descriptors were chosen, from which the final descriptors were obtained byimplementing a Fisher score filter. Secondly a Fuzzy Inference System was constructed usinglabeled training data from the descriptors, created by the user. Finally the entire study area wasclassified pixel-by-pixel by using the trained classifier and a majority filter was used to cluster theoutputs. The result was validated by visual inspection, comparison to the current map productsand by constructing Confusion Matrices, both for the training data and validation samples as wellas for the clustered- and non-clustered results.The results showed that