Generating Synthetic Remote Sensing Data with Deep Learning for Improved Wetland Classification

Wetlands in the Prairie Pothole region of Canada are critical ecosystems that support biodiversity and provide essential ecosystem services, yet they face increasing threats from agricultural expansion and climate change. Remote sensing offers a powerful tool for monitoring these landscapes over time, enabling large-scale, consistent land cover classification compared to conventional field-based methods. However, machine learning approaches often struggle with rare wetland classes, such as fens, due to limited training data. To address this challenge, a CycleGAN model, a generative adversarial network (GAN) designed for image-to-image translation, was used to generate synthetic four-band orthophoto imagery of fens from more readily available marsh imagery. The model was trained using classified wetland imagery from central Saskatchewan within the ArcGIS deep learning framework, and the resulting synthetic fen images were statistically compared to real fen images. A t-test (p < 0.05) revealed significant differences in mean pixel intensity across all bands except blue, while Jensen-Shannon divergence values (Blue: 0.1288, Green: 0.2077, Red: 0.2339, IR: 0.1885) indicated relative similarity between real and synthetic histograms. Additionally, synthetic images exhibited significantly higher mean entropy values in all four bands (p < 0.05), suggesting increased variability. These results demonstrate that CycleGAN-generated images retain key spectral characteristics of real fens while introducing additional diversity, offering a potential solution for improving wetland classification models.