Selecting Relevant Biological Variables Derived from Sentinel-2 Data for Mapping Changes from Grassland to Arable Land Using Random Forest Classifier

Abstract

Permanent grassland is one of the monitored categories of land use, land use change, and forestry (LULUCF) within the climate concept and greenhouse gas reduction policy (Regulation (EU) 2018/841). Mapping the conditions and changes of permanent grasslands is thus very important. The area of permanent grassland is strongly influenced by agricultural subsidy policies. Over the course of history, it is possible to trace different shares of permanent grassland within agricultural land and areas with significant changes from grassland to arable land. The need for monitoring permanent grassland and arable land has been growing in recent years. New subsidy policies determining farm management are beginning to affect land use, especially in countries that have joined the EU in recent waves. The large amount of freely available satellite data enables this monitoring to take place, mainly owing to data products of the Copernicus program. There are a large number of parameters (predictors) that can be calculated from satellite data, but finding the right combination is very difficult. This study presents a methodical, systematic procedure using the random forest classifier and its internal metric of mean decrease accuracy (MDA) to select the most suitable predictors to detect changes from permanent grassland to arable land. The relevance of suitable predictors takes into account the date of the satellite image, the overall accuracy of change detection, and the time required for calculations. Biological predictors, such as leaf area index (LAI), fraction absorbed photosynthetically active radiation (FAPAR), normalized difference vegetation index (NDVI), etc. were tested in the form of a time series from the Sentinel-2 satellite, and the most suitable ones were selected. FAPAR, canopy water content (CWC), and LAI seemed to be the most suitable. The proposed change detection procedure achieved a very high accuracy of more than 95% within the study site with an area of 8766 km2.

Keywords

• random forest
• change detection
• mean decrease accuracy
• grassland mapping
• biological variables
• Czechia