A Multi-Resolution GeoAI and Machine-Learning Framework for Spatially Validated Environmental Resilience Assessment Across the Chesapeake Bay Watershed in the Mid-Atlantic United States (1984-2024)

Watershed-scale environmental resilience assessments need to identify areas where multiple environmental and social vulnerabilities overlap. However, vulnerability maps often combine independently derived layers without testing whether their predictive models transfer across geographic regions. This study presents a GeoAI and machine-learning framework for spatially validated environmental resilience assessment across the Chesapeake Bay watershed.

A uniform 1 km grid comprising 181,381 cells was used to integrate publicly accessible environmental and socioeconomic datasets covering terrain, soils, land cover, climate, hydrology, water quality, flood susceptibility, vegetation dynamics, and social vulnerability. Multi-decadal vegetation dynamics from 1984–2024 were derived from Landsat imagery using Google Earth Engine. Domain-specific machine-learning models were evaluated using station-grouped and spatial-block cross-validation. Entire physiographic provinces were withheld during spatial-block validation to assess geographic transferability. Reliability criteria were applied before predictive outputs were incorporated into the multidomain assessment.

Station-grouped validation showed moderate-to-strong performance for dissolved oxygen (R² = 0.626) and nitrogen (R² = 0.338). Agriculture-linked nitrogen showed R² ≈ 0.26, while flood classification achieved an AUC of 0.79. Spatial-block validation revealed substantially weaker transferability. Pooled R² was −0.37 for nitrogen, −0.30 for phosphorus, and −2.07 for agriculture-linked nitrogen. The largest agricultural fold failure had an R² of −9.76 and coincided with shifts in observed nitrogen concentrations and soil-drainage composition between training and held-out regions, indicating environmental distribution shift. Dissolved oxygen retained a pooled spatial-block R² of 0.485, but no physiographic province met the pre-specified reliability criterion. These results informed a reliability-gated multidomain vulnerability framework. Where direct ML prediction lacked sufficient spatial transferability, nitrogen risk was represented using a mechanistic source–transport vulnerability indicator.

The resulting three-domain overlay identified 1,711 cells (0.9%) with elevated flood exposure, nitrogen source–transport vulnerability, and social vulnerability. Getis-Ord Gi* analysis identified 1,693 cells (98.9%) within statistically significant multidomain hotspots. Local Moran's I classification agreed with the Gi* classification for 99.4% of cells. A stricter evidence criterion identified 956 cells (0.5%) with three elevated domains and independently validated nitrogen evidence.

The results show that geographic transferability is critical when using machine-learning predictions for watershed-scale vulnerability mapping. The resulting priority map provides a screening framework for directing monitoring, field investigation, and resilience interventions toward areas where multiple environmental and social pressures converge. The framework uses freely accessible datasets and open analytical tools and is designed to be reproducible and transferable to other watersheds. The complete code and analytical workflow will be publicly released to support replication.