Learning to Improve Earth Observation Flight Planning

Robert A. Morris, Nikunj C. Oza, Leslie Keely, Elif Kürklü, Anthony StrawaAAAI 2008

Introduction

Modern Earth science missions, such as the Intercontinental Chemical Transport Experiment (INTEX-B), rely on a complex "sensor web" comprising satellites, ground stations, and airborne laboratories. The primary objective of these missions is to track the global movement of pollutants and greenhouse gases. However, coordinating these heterogeneous assets presents significant challenges. Scientists are often overwhelmed by massive datasets, making it difficult to identify in real-time where predictive models diverge from reality. Furthermore, manually planning flight paths that maximize scientific value while adhering to strict constraints—such as fuel limits, instrument requirements, and Special Use Airspaces (SUAs)—is an arduous and high-stakes task.

To address these issues, NASA researchers have developed an automated, closed-loop system. This system leverages machine learning to detect discrepancies between atmospheric models and satellite observations, subsequently translating these "surprises" into optimized flight plans for research aircraft.

System Architecture

The system is designed as a daily operational cycle that integrates data collection, analysis, and planning into a cohesive loop. By focusing mobile sensors on areas where current models fail, the system accelerates the refinement of our understanding of complex atmospheric processes.

The closed-loop architecture integrating data assimilation and automated flight planning.

As illustrated in the architecture, the process begins with the "Sensor Web" providing data and predictions. The Data Assimilation Assistant identifies interesting targets, which are then processed by the Flight Planning Assistant to generate observation plans. These plans are visualized for human review before being executed by the sensors, completing the cycle.

Methodology

The system's functionality is divided into three core components:

1. Data Assimilation and Anomaly Detection

Rather than simply presenting raw data, the system identifies the "delta" between model predictions and actual observations. It utilizes the MOZART chemical transport model for predictions and compares them against satellite data from instruments like AIRS and MOPITT. To detect anomalies that deviate from standard atmospheric profiles, the researchers employed One-class Support Vector Machines (SVMs) and K-means clustering. Waypoints are prioritized based on a mathematical definition of "interestingness": a location is high-priority if there is a significant conflict between the satellite's measurement and the model's prediction.

2. Automated Flight Planning

The system treats flight planning as an Orienteering Problem (OP). Given a set of high-value waypoints, the goal is to determine a path that maximizes the total scientific "score" within a fixed time budget. The system uses a greedy stochastic algorithm known as Heuristic Biased Stochastic Sampling (HBSS). This approach allows the planner to efficiently explore the search space and suggest multiple high-value routes that respect physical aircraft constraints and restricted airspaces.

3. Visualization and Human-in-the-Loop

To ensure scientific oversight, the system exports data to a 3D geo-registered environment using a tool called Mercator (integrated with Google Earth). This allows mission planners to overlay flight legs, satellite tracks, and predicted pollution plumes within a single spatial context, facilitating rapid validation of the machine-generated plans.

Results and Discussion

The system was validated using data from the 2006 INTEX-B mission. Key performance metrics and qualitative observations include:

  • Scalability and Speed: The system successfully processed over 130,000 satellite data points. For a typical problem involving 50 candidate waypoints, the planner generated a feasible, optimized flight plan in approximately three minutes.
  • Targeting Accuracy: Qualitative analysis using the visualization tool confirmed that the ML-driven waypoints correctly targeted significant phenomena, such as high Carbon Monoxide (CO) plumes over Mexico City.
Visualizing optimized flight paths and waypoints in a 3D geo-registered environment.

As shown in the visualization, the system can map complex routes (white lines) through various observation points (O) and waypoints (W) while accounting for geographic features and potential hazards. This spatial context is vital for scientists to trust and refine the automated suggestions.

Conclusion and Future Work

While the system demonstrates a significant leap in automating Earth observation logistics, several areas for improvement remain. One primary challenge is interpretability; while SVMs are effective at finding anomalies, the mathematical distance from a hyperplane does not always translate to a physical concept that scientists can easily intuit.

Future work will focus on more rigorous, quantitative comparisons between human-planned and machine-planned missions to fully assess the utility of the automated approach. Additionally, further research into parameter tuning is required to ensure the system consistently generates a manageable number of high-priority waypoints for the planning algorithm to process effectively.

Summary updated

BibTeX
@inproceedings{morris2008learning,
  author = {Robert A. Morris and Nikunj C. Oza and Leslie Keely and Elif Kürklü and Anthony Strawa},
  title = {{Learning to Improve Earth Observation Flight Planning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2008},
}
Figure