Estimating the Impact of COVID-19 on Travel Demand in Houston Area Using Deep Learning and Satellite Imagery

1. Problem Statement

The paper addresses the challenge of accurately estimating travel demand and economic activity during periods of significant disruption, specifically the COVID-19 pandemic. Traditional data collection methods (e.g., traffic sensors, surveys) often suffer from latency, high costs, or incomplete coverage. The authors propose using high-resolution satellite imagery combined with deep learning to provide a cost-effective, scalable, and objective method for monitoring transportation infrastructure usage and inferring economic trends in the Houston metropolitan area. The core objective is to quantify the reduction in vehicle presence at key economic hubs (universities, malls, community centers) before and during the pandemic.

2. Methodology

The study employs a pipeline integrating remote sensing data acquisition, deep learning-based object detection, and statistical analysis.

A. Data Acquisition

• Source: High-resolution satellite imagery was acquired via Google Earth Engine and Google Earth Pro.
• Resolution: The images have a Ground Sample Distance (GSD) of approximately 15 cm to 30 cm, allowing for the detection of individual vehicles.
• Study Area: Eight specific locations in Houston, Texas, were selected to represent diverse economic sectors:
  • Community/Commercial: Evelyn Rubenstein JCC, Braeswood Square, Hillcroft/S. Braeswood intersection.
  • Education: University of Houston (UH) Parking Lots.
  • Major Retail: Meyerland Plaza, Katy Mills, Houston Chinatown.
• Temporal Scope: Images were collected from 2019 (pre-pandemic) and 2020 (during the pandemic). A total of 127 images were collected across these sites.

B. Deep Learning Model

• Algorithm: The authors utilized the Detectron2 framework with the Faster R-CNN (Region-based Convolutional Neural Network) architecture for object detection.
• Training Dataset: The model was trained on the Cars Overhead with Context (COWC) dataset, which contains ~33,000 annotated cars from six global locations (Toronto, Selwyn, Potsdam, Vaihingen, Columbus, Utah).
  • Preprocessing: Only 3-band RGB images were used (grayscale datasets were excluded).
  • Split: 70% training, 30% testing.
  • Training Parameters: The model was trained for 9,000 epochs on a Google Colab Pro instance.
• Inference Process:
  1. Collected satellite images were cropped into 256×256 pixel tiles.
  2. Tiles were passed through the trained Detectron2 pipeline.
  3. Non-Maximum Suppression (NMS): Overlapping bounding boxes from adjacent tiles were discarded to prevent double-counting vehicles in tile boundaries.

C. Analysis

The detected vehicle counts for 2020 were compared against 2019 baseline data to calculate percentage changes in parking density, serving as a proxy for travel demand and economic activity.

3. Key Contributions

• Novel Application of Deep Learning: Demonstrates the efficacy of combining Detectron2/Faster R-CNN with high-resolution satellite imagery for specific travel demand estimation in a metropolitan context.
• Multi-Sector Economic Monitoring: Unlike studies focusing solely on road networks, this research targets specific "nodes" of economic activity (universities, shopping malls, community plazas) to provide granular insights into how different sectors were impacted.
• Scalable Framework: Proposes a methodology that can be replicated for other cities or disaster scenarios (e.g., hurricanes) without relying on ground-based sensor infrastructure.
• Validation of Satellite Data: Confirms that satellite-derived indicators can serve as reliable complements to traditional socio-economic assessment frameworks.

4. Results

• Model Performance: The trained vehicle detection model achieved an average accuracy of 90% on the test set.
• Impact Quantification:
  • The study observed a visible and quantifiable reduction in vehicle presence across all monitored locations during the pandemic.
  • Average Reduction: The number of cars detected at the selected locations decreased by an average of 30% in 2020 compared to 2019.
• Visual Evidence: The paper provides comparative imagery (Figures 3–9) showing stark differences in parking lot occupancy between the two years, particularly at the University of Houston and major shopping centers like Katy Mills.

5. Significance and Limitations

Significance:

• Decision Support: The findings offer transportation agencies and policymakers a reliable, remote tool for real-time or near-real-time monitoring of economic recovery and infrastructure usage during crises.
• Cost-Efficiency: Reduces the need for expensive, labor-intensive ground inspections or vehicle-based surveys.
• Disaster Resilience: The methodology is applicable beyond pandemics, suitable for monitoring impacts of natural disasters (e.g., hurricanes) on local economies.

Limitations & Future Work:

• Temporal Resolution: The reliance on Google Earth Pro limits the frequency of available images (often only a few per year), introducing uncertainty in daily or weekly trend analysis. The authors suggest integrating commercial satellite vendors for higher temporal frequency.
• Data Integration: To improve economic impact estimation, the authors recommend fusing satellite data with other sources such as employment statistics, surveys, and price indices.
• Generalization: While the model performed well, future work should address the variability of building types and geographical regions to ensure robust generalization.

In conclusion, the paper successfully validates that high-resolution satellite imagery, processed through advanced deep learning models, is a powerful tool for quantifying the macro- and micro-economic impacts of global events like the COVID-19 pandemic on travel demand.