Open Data Science Europe workshop 2022

Lucie Stará

PhD student at Department of Geomatics, CTU in Prague


Sessions

06-15
12:50
20min
Random forest classification based on selected CORINE land cover classes and Sentinel-2 data
Lucie Stará

This study investigates image classification based on open data and the usage of open source tools. The classification was performed on Sentinel-2 data with the use of the CORINE database. Three CORINE land use classes (permanently irrigated arable land, pastures, and natural grassland) report similar spectral responses which make it challenging to separate. Therefore, a multitemporal and multispectral approach was adopted using Sentinel-2 satellite imagery in combination with the NDVI vegetation index, Haralick’s textural measures, and topographic information. The workflow identifies a methodology for combining various groups of input data (optical, NDVI, textural, topographic) and explores the suitable use of the Random Forest classifier for the task. Initial image processing was performed in QGIS software, the classification was implemented in the Spyder environment. The classification was carried out in three different European locations. The results present a strong separation of arable land (F1 score over 96%) from the other two classes. Pastures and natural grassland classes achieved F1 in the range of 76% to almost 85% in both cases. In conclusion, we discuss the suitability of the CORINE database for such classification problems.

Lobby - Poster session
06-16
12:50
20min
Random forest classification based on selected CORINE land cover classes and Sentinel-2 data II.
Lucie Stará

This study investigates image classification based on open data and the usage of open source tools. The classification was performed on Sentinel-2 data with the use of the CORINE database. Three CORINE land use classes (permanently irrigated arable land, pastures, and natural grassland) report similar spectral responses which make it challenging to separate. Therefore, a multitemporal and multispectral approach was adopted using Sentinel-2 satellite imagery in combination with the NDVI vegetation index, Haralick’s textural measures, and topographic information. The workflow identifies a methodology for combining various groups of input data (optical, NDVI, textural, topographic) and explores the suitable use of the Random Forest classifier for the task. Initial image processing was performed in QGIS software, the classification was implemented in the Spyder environment. The classification was carried out in three different European locations. The results present a strong separation of arable land (F1 score over 96%) from the other two classes. Pastures and natural grassland classes achieved F1 in the range of 76% to almost 85% in both cases. In conclusion, we discuss the suitability of the CORINE database for such classification problems.

Lobby - Poster session