Modelling of the above-ground biomass and ecological composition of semi-natural grasslands on the strenght of remote sensing data and machine learning algorithms

Abstract

Semi-natural grasslands represent a vital source of forage and fodder for livestock farming, fulfilling an essential ecological function.This study upholded the efficacy of Sentinel-2 time series data in conjunction with ground data on aboveground biomass (AGB), ecological composition, and floristic composition for the development of a prediction model to determine AGB yield and species distribution in a semi-natural grassland. Two machine learning (ML) approaches, namely K-means clustering, and Random Forest (RF) classification were employed in a semi-natural grassland in Southern Europe. The study was subjected to a qualitative ground survey that identified two main groups. The first group (Cluster 1) was composed of nitrophilous species or nitrophilous species combined with xerophilous species and some hygrophilous species.The second group, on the other hand, consisted of only hygrophilous species or hygrophilous species associated with Ranunculus acris, Bromus hordeaceus, Brassica nigra, Lolium perenne and Trifolium sp. (Cluster 2).Using the above-mentioned algorithms, the study identified thetwo distinctive areas based on AGB estimates in the period beginning in March (vegetative growth) and extending to the first half of May. The R2 value for K-means classification was 0.59, with a root mean square error (RMSE) value of 1 t ha−1.The R2 value for the RF classification was 0.69, with an RMSE value of 0.59 t ha−1. The integration of ground-based data and ML classification techniques enables the generation of accurate, dynamic information on grassland, core attributes that play an integral role in sustainable ecosystem management and conservation.