This thesis addresses the use of superpixels as an intermediate unit for the classification of crop types in remote sensing. Multispectral images obtained from the satellite Sentinel-2 are segmented into superpixels using the SLIC and Quickshift methods. Then, over the average time series of individual superpixels, classification is carried out using the LightGBM and SITS-BERT methods. On the 2017 data on crop types in Slovenia, the superpixel approach is compared with the classification of data for whole parcels and individual pixels. The use of superpixels has proved to be particularly sensible in cases where a rougher description of the area is sufficient for classification, as such segmentation often aggregates data for small or narrow parcels. In the case of larger parcels, however, superpixels prove to be an effective summarisation technique that facilitates the interpretation of satellite data without the use of parcel data.
|
