Mowing Event Temporal Localization on Dense Satellite Time Series Using Foundational Models

Abstract. Mowing event temporal localization in dense satellite time series is essential for agricultural monitoring and Common Agricultural Policy (CAP) compliance. In this study, we investigate the use of Foundation Models (FM) as frozen encoders which construct useful features for precise temporal mowing event localization. We study the Prithvi-EO-2.0 family of FM, at different encoder sizes, paired with a lightweight MLP localization head, and compare against a 3D CNN baseline. Input timeseries from the Harmonized Landsat Sentinel-2 (HLS) dataset, are comprised of observations across six common spectral bands over the growing season in Central Greece. Moreover, a sliding window approach is used to process the input data adding tunable hyperparameters to the architecture. We further contribute 451 newly annotated parcels with Day of Year(DOY) labels for each event that occurred, extending an existing benchmark for standardized mowing event monitoring evaluation. Our results reveal a connection between event detection and temporal localization. The 3D CNN baseline achieves the highest F1 score of 83.63%, while the Prithvi-EO-2.0 600M variant attains the best Mean Temporal Error (MTE) of 0.83 timesteps, representing a 33% reduction compared to the baseline model. On top, we conduct an ablation study across model sizes, window sizes, and tolerance values which shows that scaling in a certain area of model size yields the largest performance gain. Additionally, window size and tolerance of detection are important for performance across all model sizes. The dataset and code are available at the official repo.