Grassland cut detection based on Sentinel-2 time series to respond to the environmental and technical challenges of the Austrian fodder production for livestock feeding

Abstract

The relationship between yield and quality of grassland fodder is an important factor for livestock farmers to consider when deciding the timing of grassland cuts. Even for experienced farmers, estimating this relationship is difficult because external factors such as weather can affect the growth rate of plants. Satellite remote sensing techniques have good potential for the modelling of fodder biomass and quality. However, more accurate grassland cut dates are needed. Unfortunately, current cut detection methods vary in both form, function and accuracy and a common framework for cut detection does not yet exist, especially for in-season cut estimation. The aim of this study is to develop an improved methodology for Austrian fodder production systems that can be used for both post-season and in-season cut detection. In the first case, cuts are detected at the end of the vegetation period and used for post-season assessment at national and regional levels. In the second case, cuts are detected in near-real-time during the season, and they can be used to model yield and quality. The approach is tested across the Austrian landscape on mown grasslands using freely available Copernicus Sentinel-2 imagery on the Google Earth Engine (GEE). The method uses a fitted NDVI curve obtained by iteratively smoothing the upper envelope of the actual observations using the Whittaker smoother. Cuts are detected by meeting a threshold in the difference between the fitted NDVI curve and the actual observed NDVI values. Additional innovative steps were developed to reduce the impact of cloud cover on the results for timely and accurate cut detection, including a gradient boosting algorithm to detect false positive cuts due to missed clouds. A set of rules for the method was trained on 97 plots and validated on a further 502 plots over three years (2020, 2021 and 2022). The method performed well in plots with two to four cuts per season, but occasionally struggled in more intensively managed systems, most likely due to a rule preventing two cuts within 27 days. A pooled post-season cut date detection f-score of 0.79 suggests that the method can accurately detect cuts, with absolute detection errors of <5 days. The main misclassifications were the results of (thin and semi-transparent) clouds being erroneously detected as cuts. With an f-score of 0.80 and total absolute delayed detection of approximately 6 days, the in-season cut date detection performs similarly to the post-season configuration. This indicates a high potential for near-real-time applications. Our results moreover indicate a general improvement of existing methodologies despite performing the cut detection across the fragmented, mountainous and diverse Austrian landscape. Future work will include grazed pastures and natural grassland with less intensive human activity.