Possibilities of urban land cover classification based on fusion of DESIS and PlanetScope satellite image time-series

Abstract

The use of hyperspectral remote sensing data allows for the identification of more detailed land cover types in urban environments
than commonly available multispectral satellite imagery. A limitation of hyperspectral satellite imagery, compared to multispectral, is
its currently lower spatial resolution. This study combines high-resolution fused data from multispectral PlanetScope and hyperspectral
DESIS satellite imagery to identify diverse urban surfaces using Random Forest machine learning classifier. The data were prepared
for a sample area in Budapest, captured at four different date: June 24, August 18, and October 10 of 2022, and February 15, 2023.
Classification of main land cover groups—impervious surface, vegetation, and soil—achieved nearly 90% overall accuracy.
Additionally, we successfully differentiated subcategories with acceptable accuracy, such as metal, concrete, asphalt, clay roof tiles,
and flat roofs. The classification accuracy was further improved by incorporating external vector data, specifically OpenStreetMap
building polygons. From the classified time-series data, we identified permanent and variable pixels within the sample area for the
study period.