Enhancing grassland cut detection using Sentinel-2 time series through integration of Sentinel-1 SAR and weather data

Abstract

The detection of grassland cuts is relevant for modelling grassland yield and quality because information on cut dates and cut intensity aids in the modelling of the nutrient biomass ratio of fodder. This research improves an existing grassland cut detection methodology developed for Austria based on Sentinel-2 (S2) optical time series. To further improve the detection accuracy, the new method incorporates Sentinel-1 (S1) Synthetic Aperture Radar (SAR) and daily weather data utilizing a machine learning-based model (Catboost). Cuts are first identified through a threshold-based comparison between a fitted idealized grassland growth curve and the observed NDVI values. The Catboost model subsequently addresses limitations in S2 data caused by cloud cover and other sub-optimum observation conditions. The Catboost model (1) identifies missing cuts in periods with no S2 data, and (2) eliminates false positive cuts. Weather data is utilized to identify the start of the cutting season and to define the (minimum required) time span between two consecutive cuts. Results demonstrate an improvement in cut date f-score (from 0.77 to 0.81), a reduced false detection rate (from 0.21 to 0.16), and a slight decrease in mean absolute error between true and estimated cut dates (from 4.6 to 4.1). The improvement in the accuracy was more evident for plots with high mowing frequency, while some remaining false detections were evident for extensively managed grasslands. The incorporation of S1 SAR and weather data enables the cut detection for the entire calendar year and eliminates the need for fixed growing season start/end dates. However, S1 SAR data alone did not provide reliable detection accuracy, showing its limitations in depicting vegetation dynamics for grassland. Overall, the improvements in accuracy and flexibility demonstrate the efficacy of the enhanced methodology, emphasizing the potential of combining S1 and S2 with weather data in large scale and cost-efficient grassland monitoring.