Detection of land cover changes based on the Sentinel-2 multitemporal data on the GEE platform

The world around us is constantly changing, and humans contribute to many of these changes. Land cover and land use (LCLU) changes over time have a significant impact on the functioning of the Earth, particularly climate change and global warming. Spatial data of LCLU changes find important applications in land management, monitoring the sustainable development of agriculture, forestry, rural areas, assessing the state of biodiversity and urban planning.

In the frame of the InCoNaDa project "Enhancing the user uptake of Land Cover / Land Use information derived from the integration of Copernicus services and national databases”, the maps of land cover (LC) changes were developed for two study areas - the Łódź Voivodeship in Poland and the Viken County in Norway. The detection of LC changes was performed on the annual bases for the period 2018-2021 based on the analysis of multitemporal optical data from the Sentinel-2 mission. The Google Earth Engine (GEE) platform was used, which allows to analyze satellite data and to perform spatial analyses anywhere in the World while providing computing power. The LC change detection method was divided into two phases. The first phase is based on the analysis of spectral signatures, and the second phase applies the machine learning Random Forest algorithm. The classification was performed separately for each time interval: 2018-2019, 2019-2020, 2020-2021. In this way, three independent classification models were developed for each study area. The following three LC change classes were distinguished:  a) no-change, b) forest loss, and c) construction sites and newly built-up areas. The minimum mapping unit (MMU) was 0.2 ha. The LC change detection models reached high accuracy - in both study areas for all time intervals, the overall accuracy was equal to or greater than 0.97 and the Kappa coefficient than 0.95. The independent verification carried out based on the aerial orthophotos proved that the overall accuracy of the LC changes is pretty good for both study areas (around 0.9). The changes occurring in the construction sites and newly built-up area class reached slightly lower accuracy and has the lowest precision. The presented method showed its universality and adaptability, giving the possibility for further development. We will present the method, algorithm, results and their verification for Poland and Norway.