Estimation of Agricultural Intensification Through Semi-Supervised Deep Learning on Sentinel-2 Imagery

This study presents a semi-supervised approach for employing deep learning change detection models to identify vegetation clearance events in arable cropping systems, using bitemporal Sentinel-2 imagery. The resulting change maps are aggregated to estimate the frequency of relevant farming activities (i.e., harvesting, plowing, chemical desiccation, etc.), serving as a proxy for quantifying agricultural intensification. To address the extensive training data requirements of the model architectures, we generate pseudo-labels based on rule-based criteria applied to vegetation and soil spectral indices, identifying instances of above-ground biomass removal between image pairs. We explore two training strategies, including fine-tuning pre-trained models on a benchmark dataset and training models from scratch, applied in two agricultural regions of central Greece (Larissa and Viotia). Model performance is validated against in-situ and photointerpretation data, demonstrating robust performance and a detection rate exceeding 85% for harvest events. The most accurate models are then deployed to estimate agricultural intensification over the period 2021–2023. Results reveal high spatial variability in the number of detected changes, indicating differences in land management intensity depending on the region, cultivation year, and crop type. This cost-efficient framework enables scalable monitoring of land use dynamics, offering insights into farmers' activity patterns and supporting the development of sustainable land management recommendations.