| Popis: |
Namen diplomske naloge je izboljšati ločljivost podatkov Sentinel-2. V nalogi se preizkusi metoda super-ločljivosti, ki poteka na podlagi strojnega učenja dveh konvolucijskih nevronskih omrežij. Podatki Sentinel-2 so glede na posamezne spektralne kanale podani v prostorski ločljivosti 10 m, 20 m ter 60 m. Cilj naloge je pridobiti celoten nabor podatkov v ločljivosti 10 m. Opisane so spletne aplikacije Google Colab ter osnove strojnega učenja s poudarkom na nevronskih omrežjih. Predstavljena je metoda izboljšave ločljivosti DSen2. Nevronska omrežja so dodatno naučena z osmimi posnetki območja Slovenije. Opisan je postopek izbora posnetkov ter postopek izboljšave ločljivosti. Na koncu naloge je narejena primerjava obdelanega posnetka s prvotnim. The purpose of this thesis is resolution enhancement of Sentinel-2 data. The thesis tests a super resolution method based on machine learning of two convolutional neural networks. Sentinel-2 data are given in spatial resolutions of 10 m, 20 m and 60 m, depending on the individual spectral channels. The aim of the thesis is to obtain the complete dataset in 10 m. Web application Google Colab and the basics machine learning focused on neural networks are described. A convolutional neural network called DSen2 is presented. The network has been enhanced with eight different Sentinel-2 images of the area of Slovenia. Describing the selection of images, and the process of resolution enhancement. At the end of the thesis a comparison of the results and the original data are described. |
