Machine learning for grazing event detection and pasture utilisation quantification from sentinel-2 data

Effective pasture management in pasture-based dairy systems (P-BDS) requires accurate knowledge of when and where grazing occurs and how much biomass is consumed, yet manual recording methods are time consuming and prone to error. This study compared machine learning approaches for automated grazing event detection and pasture utilisation quantification from both rising plate meter and interpolated Sentinel-2 satellite data. Performance was evaluated against GPS-tracked cow grazing records across 12 commercial dairy farms in New South Wales, Australia (July 2022–June 2024). Eleven approaches were evaluated across within-year and cross-year validation scenarios to assess temporal transferability. Random Forest achieved optimal within-year detection performance (F1 score = 0.878, precision = 0.938, recall = 0.825), whilst One-Class Support Vector Machine (OCSVM) demonstrated superior cross-year transferability (F1 score = 0.692), outperforming supervised models by 7.6% on independent Year 2 data despite supervised models experiencing average performance degradation of 24.2%. Farm-level detection variability ranged from F1 = 0.500 to 0.815, with site-specific factors exerting stronger influence than regional characteristics. Independent pasture utilisation agreement validation on 218 GPS-confirmed events demonstrated strong biomass quantification concordance between satellite and ground measurements (pre-grazing biomass R² = 0.966; post-grazing biomass R² = 0.998; biomass removal R² = 0.922), establishing proof-of-concept that satellite systems can both identify when grazing occurs and accurately quantify biomass changes. Temporal alignment constraints between satellite revisit schedules and ground measurements limited validation to 19.4% of total events, highlighting the need for daily or near-daily observations through multi-sensor fusion approaches to achieve deployment in commercial P-BDS.