Arabi et al. (2026) Satellite-based condition assessment of urban green stormwater infrastructure using a hybrid machine learning framework

Identification

• Journal: Scientific Reports
• Year: 2026
• Date: 2026-09-22
• Authors: Shiva Arabi, Virginia B. Smith, Peleg Kremer, Bridget Wadzuk, Xun Jiao
• DOI: 10.1038/s41598-026-67716-2

Research Groups

• Department of Civil and Environmental Engineering, Villanova University, Villanova, PA, USA
• Department of Geography and the Environment, Villanova University, Villanova, PA, USA
• Department of Electrical and Computer Engineering, Villanova University, Villanova, PA, USA

Short Summary

This study developed a hybrid machine learning framework integrating high-resolution multispectral satellite imagery, unsupervised anomaly detection, and supervised classification to automate the condition assessment of urban green stormwater infrastructure (GSI). The framework successfully evaluated multiple GSI condition indicators, demonstrating strong performance for vegetation stability and providing a scalable tool for maintenance prioritization.

Objective

• To evaluate the potential of high-resolution multispectral satellite imagery combined with a hybrid machine learning framework for automated condition assessment of urban green stormwater infrastructure (GSI) sites across multiple indicators (overall performance, vegetation health, vegetation stability, erosion, and debris presence).
• To develop a scalable, data-driven monitoring approach to support city-wide GSI maintenance planning and improve long-term effectiveness.

Study Configuration

• Spatial Scale: Philadelphia, Pennsylvania, USA. The study included 395 GSI sites (295 basins and 100 bioinfiltration facilities). Satellite imagery had a spatial resolution of approximately 3 meters.
• Temporal Scale: Near-daily satellite imagery revisit frequency. Imagery was temporally matched to inspection records, prioritizing same-day or within one day of inspection. The dataset had limited representation of winter inspections.

Methodology and Data

• Models used:
  • Hybrid One-class Support Vector Machine (OCSVM) for unsupervised anomaly detection.
  • Extreme Gradient Boosting (XGBoost) for supervised classification.
  • Synthetic Minority Oversampling Technique (SMOTE) for addressing class imbalance.
  • SHapley Additive exPlanations (SHAP) for model interpretability.
• Data sources:
  • Satellite imagery: High-resolution PlanetScope imagery (Planet Labs), providing ~3 meter spatial resolution with four spectral bands (blue, green, red, near-infrared). Ortho Scene Surface Reflectance products were used.
  • Observation data: Philadelphia Water Department (PWD) inspection data, including georeferenced polygon footprints for GSI sites, installation and inspection dates, GSI type, polygon area, perimeter length, and a four-level ordinal performance rating scale for five condition indicators (Overall GSI Performance, Vegetation Health in Basin Area, Vegetation Stability, Erosion within Basin Area, and Debris or Trash within Drainage Area).
  • Environmental data: Total daily rainfall accumulated over the three days preceding each inspection, obtained from the Iowa Environmental Mesonet (IEM) ASOS network for the PHL station.
  • Derived features: Six remote sensing indices (Normalized Difference Vegetation Index (NDVI), Soil-adjusted Vegetation Index (SAVI), Green–red Vegetation Index (GRVI), Normalized Difference Water Index (NDWI), Brightness Index (BI), and Modified Normalized Difference Built-up Index (NDBI)), spectral band reflectance statistics, hydrological variables, geometric attributes (area, perimeter), and temporal variables (asset age, inspection day of year).

Main Results

• The hybrid OCSVM–XGBoost framework demonstrated the ability to assess GSI conditions across all five inspection components.
• Vegetation Stability showed the strongest performance (accuracy = 0.93; F1-score = 0.93), with the lowest Brier score (0.06).
• Debris or Trash within Drainage Area achieved an accuracy of 0.78 and an F1-score of 0.77.
• Vegetation Health in Basin Area achieved an accuracy of 0.72 and an F1-score of 0.77.
• Erosion within Basin Area achieved an accuracy of 0.70 and an F1-score of 0.73.
• Overall GSI Performance was the most challenging target (accuracy = 0.61; F1-score = 0.59), with the highest Brier score (0.27).
• AUC-ROC values ranged from 0.57 to 0.66 across all targets, indicating modest overall discrimination.
• Compliant conditions were generally classified more reliably than non-compliant conditions across most targets, except for Overall GSI Performance, where the non-compliant F1-score (0.70) exceeded the compliant F1-score (0.41) due to a more balanced class distribution.
• Risk score distributions showed clear and statistically significant separation between compliant and non-compliant cases across all targets (p < 0.001).
• Feature importance analysis revealed that the OCSVM anomaly score, geometric characteristics (perimeter length, GSI type, coverage area, age), recent rainfall, and various spectral vegetation indices (GRVI, SAVI, BI, NIR reflectance, blue reflectance) were key predictors depending on the inspection component.
• Ablation analysis confirmed that SMOTE-based class balancing substantially improved non-compliant detection for imbalanced targets, and the OCSVM anomaly score provided further statistically significant improvement for Debris or Trash within Drainage Area (p = 0.023) and improved discrimination for Vegetation Stability (ROC-AUC increased from 0.61 to 0.66).

Contributions

• Developed and validated a novel hybrid machine learning framework (OCSVM–XGBoost) for automated, scalable condition assessment of urban GSI using high-resolution multispectral satellite imagery.
• Extended GSI remote sensing assessment beyond vegetation characteristics to include comprehensive functional indicators like erosion, debris presence, and overall performance, which are critical for municipal maintenance decisions.
• Demonstrated a practical approach to address class imbalance in GSI monitoring datasets by integrating unsupervised anomaly detection (OCSVM) and synthetic minority oversampling (SMOTE) with supervised classification (XGBoost).
• Provided a scalable and cost-effective decision-support tool for municipalities to prioritize GSI sites for maintenance, improving efficiency compared to traditional labor-intensive field inspections.
• Offered insights into the interpretability of GSI condition assessment models by identifying key spectral, environmental, and geometric features influencing predictions.

Funding

• National Science Foundation under Award No. 2426951

Citation

@article{Arabi2026Satellitebased,
  author = {Arabi, Shiva and Smith, Virginia B. and Kremer, Peleg and Wadzuk, Bridget and Jiao, Xun},
  title = {Satellite-based condition assessment of urban green stormwater infrastructure using a hybrid machine learning framework},
  journal = {Scientific Reports},
  year = {2026},
  doi = {10.1038/s41598-026-67716-2},
  url = {https://doi.org/10.1038/s41598-026-67716-2}
}