Martins Pukitis, Ints Mednieks. Classification of satellite images using Dynland technology. Ninth International Conference on Remote Sensing and Geoinformation of the Environment (RSCy2023), 127860(), Proc. SPIE 12786, 2023.
Bibtex citation:
Bibtex citation:
@inproceedings{15363_2023,
author = {Martins Pukitis and Ints Mednieks},
title = {Classification of satellite images using Dynland technology},
journal = {Ninth International Conference on Remote Sensing and Geoinformation of the Environment (RSCy2023)},
volume = {127860},
publisher = {Proc. SPIE 12786},
year = {2023}
}
author = {Martins Pukitis and Ints Mednieks},
title = {Classification of satellite images using Dynland technology},
journal = {Ninth International Conference on Remote Sensing and Geoinformation of the Environment (RSCy2023)},
volume = {127860},
publisher = {Proc. SPIE 12786},
year = {2023}
}
Abstract: Satellite images are widely used for land cover classification into predefined categories. To perform it in a practicable way, field data are needed to train and test a classifier. The collection of accurate field data requires significant resources, and image data of a category may have sophisticated distribution in the feature space. “Dynland” technology was developed to overcome these problems and offer a robust and practical way to classify such images. It is based on a nonparametric clustering method applicable to data with different distributions with a following assignment of classes to clusters based on their overlapping with available reference data which can be scarce and imprecise. Dynland clustering algorithm provides meaningful clusters but requires huge computational resources currently limiting the size of processed images. To overcome this problem, we propose to apply a sampling procedure i.e. use each 𝑛-th pixel on both axes for clustering and distribute the remaining pixels to created clusters based on nearest neighbors’ search. However, such an approach reduces the classification accuracy; the analysis results of this reduction will be presented. Another way how to deal with images of large areas is to split them into fragments. To classify a fragment, reference data for all categories of interest should be available for the area it represents. This limitation should be resolved so that the available reference data are used for areas not limited to one fragment. To deal with that, a method is proposed allowing to exploit the classification result of a neighboring fragment for which the reference data were available in full. The proposed solutions widened the applicability of the Dynland technology e.g. enabled the production of a classified map of mires and peatlands for the whole territory of Latvia.