Dynamic Targeting of Satellite Observations Using Supplemental Geostationary Satellite Data and Hierarchical Planning

This paper proposes a hierarchical planning approach that integrates supplemental geostationary satellite data to extend lookahead horizons for Dynamic Targeting missions, demonstrating up to a 41% performance improvement over traditional onboard-only planners, particularly in scenarios with sparsely distributed targets.

The Gist

Imagine you are the captain of a high-speed train traveling across a vast, unpredictable landscape. Your job is to take the most beautiful, valuable photographs possible with your expensive, high-powered camera. However, you have a few strict rules:

1. You can only take 100 photos before your memory card fills up and your battery dies.
2. Your camera is heavy. It takes time to swivel and point at a new spot. You can't just snap a picture of everything instantly.
3. You only have a tiny window of vision. You have a small, cheap sensor on the front of the train that can only see about 500 meters ahead. Once you pass a spot, you can't go back.

This is the challenge faced by Dynamic Targeting (DT) satellites. They fly over Earth, trying to capture rare events like storms, wildfires, or clear views of cities, but they only know what's happening right now or a minute into the future. They often miss the best targets because they run out of time or memory before they even know the targets exist.

The New Idea: The "Weather Balloon" on the Roof

The authors of this paper propose a clever solution: Give the satellite a "long-distance view" using data from geostationary satellites.

Think of geostationary satellites as giant, stationary weather balloons hovering high above the Earth, watching the same patch of ground constantly. They can see storms forming 30 minutes away or clouds clearing over a city hours in advance.

The problem? If you try to plan your entire 100-photo trip based on a 30-minute view, the math gets incredibly complicated. It's like trying to plan a cross-country road trip while simultaneously deciding which specific gas station to stop at in the next 5 miles. The number of possibilities explodes, and the satellite's computer (which is small and slow) can't handle it.

The Solution: A Two-Step "Hierarchical" Plan

To solve this, the researchers created a two-layer planning system, like a general and a field commander working together:

1. The General (Long-Term Plan):

   • Input: Uses the "Weather Balloon" data (geostationary satellites).
   • Job: Looks at the whole journey (the next 30 minutes). It doesn't decide exactly which photo to take. Instead, it draws a blueprint. It says, "Okay, there's a huge storm coming up in 20 minutes, so we should save 40 of our 100 photos for that area. There's a clear city 10 minutes away, so let's save 20 photos for that."
   • Speed: This is a fast, simple calculation. It sets the budget for each section of the trip.

2. The Field Commander (Short-Term Plan):

   • Input: Uses the satellite's own "tiny window" sensor (the 1-minute lookahead).
   • Job: Looks at the immediate future (the next 40 seconds). It takes the General's blueprint ("Save 20 photos for the city") and figures out the exact best way to execute it. It decides, "Okay, I need to swivel left now to catch that specific building, then wait for the clouds to move, then snap the shot."
   • Speed: This is a detailed, complex calculation, but it only has to solve a small piece of the puzzle.

The Results: Why It Matters

The team tested this system in four different scenarios:

• Avoiding Clouds: Trying to find clear skies to photograph the ground.
• Population Density: Finding clear skies over crowded cities.
• Random Targets: Looking for specific, rare spots (like a volcano) scattered randomly.
• Storm Hunting: Chasing massive, fast-moving storms.

The Findings:

• For "Common" problems: When good targets are everywhere (like clear skies in a generally cloudy area), the new system is just as good as the old way. The satellite can just react quickly and find what it needs.
• For "Rare" problems: When the best targets are sparse and scattered (like a single massive storm or a specific city in a sea of clouds), the new system is a game-changer.
  • The old system (looking only 1 minute ahead) often wasted its 100 photos on "okay" targets because it didn't know a "perfect" target was coming 20 minutes later.
  • The new system (using the "Weather Balloon" data) knew to save its ammo. It held back its photos, waited for the big event, and captured it.

The Bottom Line:
By combining a "long-range map" (geostationary data) with a "close-up view" (onboard sensor), the satellite became 41% more efficient at capturing the most valuable scientific data. It's the difference between a tourist snapping random photos as they walk down a street and a professional photographer who knows exactly where the sunset will hit and waits patiently to get the perfect shot.

This approach allows future satellites to be smarter, capturing more science from rare, fleeting events without needing bigger computers or more fuel.

Technical Summary

Here is a detailed technical summary of the paper "Dynamic Targeting of Satellite Observations Using Supplemental Geostationary Satellite Data and Hierarchical Planning."

1. Problem Statement

Dynamic Targeting (DT) is a satellite mission concept where a "lookahead" sensor gathers environmental data to intelligently plan observations for a primary, high-value sensor. The goal is to maximize scientific utility (e.g., avoiding clouds or hunting storms) while adhering to strict constraints like limited onboard storage, energy, and slewing (pointing) capabilities.

The Core Challenge:
Traditional DT systems rely solely on onboard lookahead sensors, which provide a very limited "horizon" (typically ~1 minute of lookahead, or ~500 km ahead). This short-term view creates a "myopic" planning problem:

• Limited Spatial Extent: The planner cannot see the entire orbit trajectory, making it difficult to optimally distribute a limited number of observations (e.g., 100 total) across the whole flight.
• Dynamic Environments: In scenarios like storm hunting or cloud avoidance, high-value targets are often sparse and dynamic. Without long-term context, the planner may waste observations on low-value areas early in the flight, missing critical events later.
• Computational Constraints: Expanding the lookahead horizon to the full orbit increases the search space exponentially, making real-time optimal planning computationally intractable.

2. Methodology

The authors propose a solution that integrates supplemental geostationary satellite data (which provides a long-term view of the Earth) with onboard lookahead data using a hierarchical planning framework.

A. Data Sources

• Onboard Lookahead (Ground Truth): Simulated using MODIS Cloud Mask (for cloud avoidance) and IMERG precipitation data (for storm hunting). This represents the high-resolution, near-real-time data available to the satellite.
• Supplemental Geostationary Data (Extended Lookahead):
  • GOES (East/West): Provides cloud masks for the Americas (10-min cadence, 2km resolution).
  • Meteosat: Provides precipitation estimates for the Atlantic/Africa (15-min cadence, 4km resolution).
  • Note: This data has lower resolution and higher latency than onboard data but covers the entire orbit trajectory.

B. Hierarchical Planning Framework

To handle the increased complexity of a long horizon, the authors introduce a two-tier planning approach (Algorithm 1):

1. High-Level Planner (Long-Term Distribution):

   • Runs periodically (e.g., when new geostationary data arrives).
   • Divides the flight trajectory into 10-cycle partitions.
   • Distributes the total observation budget (100 observations) across these partitions based on anticipated utility derived from the geostationary data.
   • Three distribution strategies were tested:
     • Uniform: Even distribution (baseline).
     • Preinformed: Distribution based on known static targets (e.g., population density).
     • Geostationary Satellite Data-Informed (GSDI): Distribution based on the utility predicted by the geostationary data.

2. Low-Level Planner (Short-Term Execution):

   • Executes every cycle (4 seconds).
   • Uses a Beam Search algorithm (depth 10, width 3) to generate a specific action plan for the next 10 cycles.
   • Adheres to the observation budget allocated by the High-Level Planner.
   • Considers physical constraints (slewing speed, acceleration) and the high-resolution onboard lookahead data to refine the plan.

C. Simulation Setup

• Scenarios: Four problem instances were simulated:
  1. Cloud Avoidance (CA): Avoid clouds to get clear-sky images.
  2. Cloud Avoidance with Population Density (CAPD): Prioritize clear skies over populated areas.
  3. Cloud Avoidance with Random Targets (CART): Prioritize specific random static targets.
  4. Storm Hunting (SH): Prioritize areas with high precipitation rates.
• Baselines: Compared against Nadir-Only, Greedy, Greedy Historical, and an Upper Bound (omniscient) planner.

3. Key Contributions

1. Integration of Multi-Source Data: Demonstrated that supplementing onboard sensors with lower-resolution, long-range geostationary data significantly improves planning efficiency.
2. Hierarchical Architecture: Proposed a scalable planning structure that decouples long-term resource allocation (polynomial time) from short-term execution (search-based), solving the "exponential explosion" of the search space.
3. GSDI Algorithm: Developed a specific algorithm that uses geostationary data to intelligently allocate observation budgets to future flight segments.
4. Empirical Analysis: Provided a quantitative analysis of when this approach works best, identifying that it is most effective for sparsely distributed, high-utility dynamic targets.

4. Results

The study compared the Geostationary Satellite Data-Informed (GSDI) hierarchical planner against traditional planners (Uniform and Preinformed) across the four scenarios.

• Overall Performance: The GSDI planner consistently outperformed traditional DT planners.
• Scenario-Specific Gains:
  • Cloud Avoidance (CA) & Random Targets (CART): GSDI performed similarly to the best baseline planners (Uniform/Preinformed). In these cases, high-value targets are frequent enough that reactive, short-term planning is sufficient.
  • Population Density (CAPD) & Storm Hunting (SH): GSDI showed massive improvements.
    • CAPD: Outperformed the best baseline by a factor of 1.40.
    • SH: Outperformed the best baseline by a factor of 1.41.
• Key Finding: The hierarchical planner with geostationary data achieved up to 41% higher utility than traditional planners in the most challenging dynamic scenarios.
• Why it works: In CAPD and SH, high-value targets (cities, storms) are sparse clusters. Traditional planners often miss these clusters because they lack long-term context. The GSDI planner uses the geostationary data to "save" observations for the specific flight segments where these clusters are located, ensuring the satellite is ready to observe them when it arrives.

5. Significance and Future Work

• Scientific Impact: This work validates that future Earth observation missions can significantly increase their science return by leveraging existing geostationary satellite constellations (like GOES and Meteosat) to guide low-Earth orbit (LEO) satellites.
• Operational Relevance: The approach is particularly valuable for missions targeting rare, dynamic events (e.g., hurricanes, volcanic plumes, algal blooms) where missing a target due to poor planning is costly.
• Future Directions:
  • Incorporating uncertainty modeling for geostationary data.
  • Developing "Just-in-Time" commanding strategies where ground stations process geostationary data and uplink plans.
  • Extending the framework to multi-agent systems where multiple satellites collaborate.
  • Testing on realistic flight processors to ensure computational feasibility.

In conclusion, the paper demonstrates that a hierarchical planning approach utilizing supplemental geostationary data is a robust method for overcoming the limitations of short-term onboard sensing, specifically excelling in scenarios where high-value targets are sparse and dynamic.