On the Application of Remote Sensing Time Series Analysis for Land Cover Mapping: Spectral Indices for Crops Classification

Abstract

Abstract. This study aims to introduce a semi-automatic classification workflow for the production of a land use/land cover (LULC) map of the island of Sardinia (Italy) following the CORINE legend schema, and a ground spatial resolution compatible with a scale of 1:25.000. The classification is based on free high-resolution satellite imagery from Sentinel-1 and Sentinel-2 collected in 2020, ancillary data derived from Sardinian Geoportal, Joint Research Centre (JRC) and OpenStreetMap. The LULC map production includes three steps: 1) pixel-based classification, realized with two different approaches, that use i) information derived from existing thematic maps eventually re-coded in case of incoherencies observed between datasets and/or satellite data products, and ii) spectral indices and parameter thresholds defined on the basis of multitemporal analysis; 2) segmentation of Sentinel-1 and 2 annual composites, and pre-labelling of segments with the pixel-based classified map, obtaining the preliminary map; 3) visual inspection procedure in order to confirm, or re-assign, classes to polygons. The accuracy of the preliminary map was tested in a sample area and on specific class of non-irrigated crops through ground truth data collected from a detailed photo-interpretation, estimating 97% of overall accuracy. The results show a great improvement from existing thematic maps in terms of detail, with the possibility of a yearly updating of the map via automatic processes. However, some limitations were found, due to the high fragmentation of Sardinian landscape and the high variety of crop types and agricultural practices, that could affect the efficiency of the classifier.