3DUNetGSFormer: A deep learning pipeline for complex wetland mapping using generative adversarial networks and Swin transformer

A groundbreaking deep learning pipeline, 3DUNetGSFormer, achieves exceptional accuracy in wetland mapping, addressing the challenge of limited training data. This novel approach combines generative adversarial networks and Swin transformers to synthetically generate high-quality wetland training data, significantly improving classification performance. The methodology demonstrates a kappa coefficient, average accuracy, and overall accuracy of 96.99%, 97.13%, and 97.39% respectively, across pilot sites in Canada, including Gros Morne National Park. By leveraging advanced convolutional neural networks and vision transformers, this study opens new avenues for high-quality wetland data generation and classification, offering a scalable solution to ecosystem monitoring in degraded or lost wetlands.

Jamali, A., Mahdianpari, M., Brisco, B., Mao, D., Salehi, B., & Mohammadimanesh, F. (2022). 3DUNetGSFormer: A deep learning pipeline for complex wetland mapping using generative adversarial networks and Swin transformer. Ecological Informatics, 72, 101904. https://doi.org/10.1016/j.ecoinf.2022.101904