Large Scale Mowing Event Detection on Dense Time Series Data Using Deep Learning Methods and Knowledge Distillation

Abstract. The intensity of agricultural land use is a critical factor for food security and biodiversity preservation, necessitating effective and scalable monitoring techniques. This study presents a novel approach for large-scale mowing event frequency detection using dense time series data and deep learning (DL) methods. Leveraging Sentinel-2 and Landsat data, we developed a benchmark dataset of over 1,600 annotated parcels in Greece, capturing mowing events through photo-interpretation and Enhanced Vegetation Index (EVI) analysis. Four DL architectures were evaluated, including MLP, ResNet18, MLP+Transformer, and Conv+Transformer, with additional handcrafted features incorporated to assess their impact on performance. Our results demonstrate that the Conv+Transformer architecture achieved the highest improvement when enriched with additional features, while ResNet18 showed a decline in performance under similar conditions. To address data scarcity, we employed knowledge distillation, pre-training models on pseudo-labeled data derived from a dataset in Germany. This process significantly enhanced model performance, with fine-tuned ResNet18 and Conv+Transformer architectures achieving significant performance improvements. This study highlights the importance of architecture selection, feature engineering, and pre-training strategies in time series classification for agricultural monitoring. The proposed methods provide a scalable, non-invasive solution for monitoring mowing events, supporting sustainable land management and compliance with agricultural policies. Future work will explore multimodal data integration and advanced training techniques to further enhance detection accuracy.