Assessing the Potential of Sentinel-1 Terrain-Flattened Gamma Time Series for Grassland Cut Detection in Austria

Abstract

Grassland covers about one quarter of the Earth's surface. In Austria, grassland even represents the most important land-use system with 1.34 million hectares. From an agricultural point of view, its main function is the supply of forage for livestock. The precise knowledge of cut dates is an important prerequisite for biomass estimation and essential to improve grassland management. Retrieving this information on a large scale using earth observation comes with advantages regarding timeliness, costs, and data consistency. Radar observations provide gap-free measurements due to their ability to penetrate clouds. Especially Sentinel-1 C-band SAR observations are of high value as they produce dense time series.

This study aims to assess the potential of Sentinel-1 backscatter time series to detect grassland cuts in Austria. In addition, the performance of sigma naught and terrain flattened gamma naught for grassland detection is compared. In the course of this project, reference data for grassland fields over several seasons is collected via a mobile phone app provided to partner farmers. In addition to the precise cut dates, farmers also take pictures of the grasslands and provide grass height measurements and information on species distribution within the fields. Sentinel-1 sigma naught and terrain flattened gamma naught time series for VV, VH polarization, and their cross-ratio (CR) are derived for around 250 fields in total. Terrain-flattened gamma naught has shown advantages over sigma naught in mountainous areas, as it mitigates the impact of the terrain on the backscatter. This characteristic is of high importance in Austria with a significant number of steep alpine meadows. In the course of time series generation, extensive pre-processing is carried out to remove the undesired impact of radar shadows and mixed land cover, e.g., single trees or small buildings within field parcels. Using either the sigma or gamma time series and reference cut dates, we train a Gated Recurrent Unit. The designed model predicts for every input time step the probability of being a cut date. To reduce the number of false-positive cut dates we apply knowledge-based rules to this probability time series, e.g., limiting its range to the cut and growing season or defining a certain number of days in between consecutive cuts. Within the accuracy assessment, the performance of the model based on sigma naught and terrain flattened gamma naught will be evaluated and further investigated. While the designed workflow is only applied on previous seasons in a first step, it will later be tested for near-real-time cut detection.

Our preliminary results on a small scale indicate good performances for fields with distinct patterns in the time series: grassland cuts lead to a quick increase in VH polarization and CR, followed by a gradual decrease. However, some fields miss these discriminating patterns. Using a comprehensive reference database will allow a deeper understanding of the backscatter signal before and after cut events and under which conditions a cut detection fails respectively a false positive cut is detected. At a later stage, we aim to improve the result by adding meteorological data (growing degree days, rainfall, soil moisture) to the model.

In summary, this study provides new findings on the potentials and limitations of SAR based cut detection and a comparison of the performance of sigma naught and terrain flattened gamma for grassland applications in mountainous areas.