News
17.09.2024
SSL-SoilNet: Tackling the Lack of Ground Truth Data with Self-supervised Deep Learning
Our Latest publication in the IEEE Transactions on Geoscience and Remote Sensing
The challenge of insufficient ground truth data is a common obstacle in soil science due to the extensive fieldwork and lab analysis required. Our latest publication addresses this issue by implementing contrastive learning in a self-supervision context.
Our postdoctoral researcher, Dr. Nafiseh Kakhani (github.io), developed SSL-SoilNet, a pretrained model leveraging a hybrid transformer-based architecture with extensive unlabeled remote sensing imagery and climate time series data. The model was subsequently fine-tuned to predict soil organic carbon (SOC), a critical environmental parameter. Impressively, SSL-SoilNet outperformed conventional supervised learning approaches and other prevalent machine learning models, demonstrating its versatility across various applications.
Explore the full paper here: https://ieeexplore.ieee.org/document/10639449 and
access the GitHub code here: https://lnkd.in/e8H3vWe3.
Acknowledgments: Special thanks to our external collaborators from the University of Tehran: Moien Rangzan, Dr. Sara Attarchi, and Prof. Seyed Kazem Alavipanah. We also extend our gratitude to Dr. Ali Jamali, Prof. Dr. Michael Mommert, and Prof. Nikolaos Tziolas for their invaluable contributions, which significantly enriched our research, especially in the fields of computer vision and environmental science. Finally, we would like to acknowledge Thomas Klein from Neural information processing group for his meticulous review and the advanced insights he brought to this research.
We are excited about the potential impact of our work and its broader applications in the future.
Teilen
