Semi-Supervised Deep Learning for Change Detection in Agricultural Fields Using Sentinel-2 Imagery

This paper introduces an original application for detecting changes related to diverse agricultural activities through the analysis of bitemporal Sentinel-2 satellite imagery. Operating without pre-existing samples, our approach generates pseudo-labels using common rule-based Earth Observation (EO) algorithms to identify cases of abrupt loss of vegetation in pairs of consecutive cloud-free images. These artificially generated samples form the basis for training several state-of-the-art change detection (CD) methods. Evaluation on a small ground truth sample, annotated through photo-interpretation by experts, demonstrates our semi-supervised methodology’s high predictive accuracy for agricultural events detection across diverse terrains and cropping practices (i.e., mowing, grazing, harvest, plowing, stubble burning, etc.). The proposed implementation offers a cost-effective, scalable solution for real-time monitoring, providing valuable insights for agricultural activity and facilitating informed decision-making in farm management and biodiversity strategies.