Self-supervised deep learning for detection of forest disturbance types in a subtropical ecosystem using transformer and Sentinel-1 and Sentinel-2 time series data

• Published in: International journal of digital earth Vol. 18; no. 1
• Main Authors: Zhang, Ming, Li, Guiying, Lu, Dengsheng, Xu, Cong, Zhao, Haotian, Li, Dengqiu
• Format: Journal Article
• Language: English
• Published: Abingdon Taylor & Francis 31.12.2025; Taylor & Francis Ltd; Taylor & Francis Group
• Subjects: Accuracy; Clearcutting; Cloud cover; Coding; Deep learning; Forest disturbance types; Forest management; Global climate; Machine learning; Modules; Remote sensing; self-supervised deep learning; Self-supervised learning; Sentinel-1/2; Time series; Training; Transformer
• ISSN: 1753-8947; 1753-8955; 1753-8955
• DOI: 10.1080/17538947.2025.2537325

Accurate and timely detection of forest disturbance types is crucial for evaluating ecosystem health and global climate stability. Time series remote-sensing data offers valuable spatiotemporal information. However, frequent cloud cover in subtropical regions disrupts the temporal consistency of optical satellite data. In addition, the impact of different data sources on modeling accuracy and the challenge of acquiring large and labeled datasets for deep learning are considerable obstacles. In this study, a novel positional encoding module was designed to handle the irregular Sentinel-2 time series. The Transformer model based on this positional encoding module effectively fused Sentinel-1 and Sentinel-2 data. Meanwhile, self-supervised learning was used to address the issue of insufficient samples. The improved Transformer model was successfully applied to detect clear-cutting and disease disturbances with a F1 score of 0.95. Our results highlighted that appropriate encoding techniques increased the model's performance by between 5% and 14%. This research also found that while Sentinel-2 data alone yields good accuracy, combining Sentinel-1 and Sentinel-2 data improved model accuracy. The self-supervised learning model achieved higher accuracy (0.95) than the supervised model (0.86) on a limited data (20% of the training samples) and exhibited stable accuracy across different training data proportions.