Spatiotemporal Transformer Architecture for Analysis and Prediction of Seagrass in Casco Bay, Maine

Seagrass meadows are crucial coastal ecosystems that provide habitat for marine life, protect shorelines from erosion, and store large amounts of carbon. However, in recent years, they have been rapidly declining; in Casco Bay, Maine, these ecosystems are experiencing high levels of loss due to a combination of light limitation, nutrient loading, and storm disturbances. The primary objective of this research is to develop a predictive machine learning framework to determine if a given coastal seagrass area will experience loss or maintain stability based on environmental data. To achieve this, heterogeneous datasets—including bathymetry, discharge data, and parameterized storm sequences—were integrated into a unified spatiotemporal framework. In this study, we evaluated several architectures, including Logistic Regression, Random Forest, ANN, XGBoost, and a Transformer-based model. Notably, the Transformer-based architecture currently demonstrates the highest predictive performance, achieving a classification accuracy of 77.4% and an ROC-AUC of 0.846. These results were achieved despite significant data constraints and a limited feature set; satellite-derived metrics suffer from high signal noise, and standardized ground truth mappings are typically conducted at multi-year intervals, preventing the capture of immediate biological responses to environmental fluctuations. This research provides crucial information to coastal decision-makers, enabling them to target areas of high risk and implement proactive conservation measures.