Classification of Sentinel-2 satellite imagery in Iran for geological purposes using deep convolutional neural networks: a case study for soil type identification

Mostafa Kiani Shahvandi¹

1) Graduate student in geodesy and remote sensing, university of Tehran

Publication : 2nd. International Congress on science & Engineering - paris(parisconf.com)
Abstract :
In this paper, Sentinel-2 satellite imagery is used for the classification of different soil types using the method of deep convolutional neural networks. Role of the presence of different phenomena in the image, including clouds, on the accuracy of the classification is analyzed. Two case studies are presented for Iran. In the first study, a region is chosen that is approximately consisted of one main type of soil, without any other type of phenomena. It is shown that the overall classification accuracy is around 95 percent, with k=0.889 . In the second study, however, the region is mixed: there are different soil types, but there are other phenomena present as well, including large clouds. The overall accuracy of classification in this case is lower, compared to the first case, being around 92 percent and with k=0.854.
Keywords : deep convolutional neural networks Sentinel-2 soil type satellite imagery