Transformers for PlanetScope Coastal Analysis

• Aditya Bajoria, Computer Science, University of Delaware

• Xu Yuan, Computer Science, University of Delaware

Coastal ecosystems, including wetlands and sea-grass beds, provide essential ecological services but face accelerating threats from climate change, sea-level rise. Monitoring these dynamic landscapes requires high-resolution, timely data and advanced analytical techniques. In my research, I leverage Planet Scope satellite imagery, which offers near-daily global coverage at high spatial resolution, to improve the classification and mapping of coastal ecosystems. My work centers on developing and refining deep learning models, with a particular focus on implementing the “Swin Transformer” architecture for image classification tasks. To streamline model development and experimentation, I integrated Hugging Face frameworks, enabling more efficient training, fine-tuning, and evaluation of transformer-based approaches. An important part of this research involved addressing class imbalance in the dataset, as certain ecosystem types were underrepresented. I implemented data augmentation techniques and adjusted loss functions to mitigate this imbalance, ensuring that the models did not over-fit to dominant classes. Additionally, I conducted extensive hyper-parameter tuning—optimizing tolerance, percentage, and weight decay parameters—to improve overall model performance and enhance class-specific prediction accuracy. This research emphasizes reducing noise in satellite imagery and improving the differentiation among wetlands, sea-grass beds, and surrounding vegetation. The outcomes of this research are intended to advance ecological monitoring, providing insights into land loss, and to predict the labels of unlabeled images. Ultimately, this integration of high-resolution satellite data with state-of-the-art computational techniques contributes to the development of tools for data-driven decision-making in environmental management, conservation, and coastal resilience planning. This research demonstrates the potential of combining Planet Scope imagery with transformer-based deep learning to deliver timely, actionable insights for safeguarding critical coastal habitats.