project image or logo
An illustration of the adaptive sensing concept.


The large spatial scales and quick evolution of complex Earth science phenomena, such as extreme weather, often makes comprehensive sensing prohibitively difficult. Current systems for remote sensing use simple “if-then” triggers or blind observations where the model of the phenomena has no impact. Adaptive sensing can improve on these approaches by utilizing online analysis of the target phenomena to direct sensing assets to collect the most advantageous observations. This approach can focus limited resources and observational capacity to extract the maximum quality science data. The above figure illustrates this concept. Observation utility is calculated based on a model of the phenomena and some scientific goal, this utility is used to plan observations for a set of assets, the plan is executed and the collected data is assimilated back into the model. This cycle then repeats while the target phenomena is active. The Planning Observations for Intelligent Science Experimentation (POISE) project developed and performed simulation testing of a prototype adaptive sensing system for collecting observations of hurricanes with the scientific objective of minimizing uncertainty in forecasted sea level pressure.


When studying these complex phenomena, it is often the case that a set of heterogeneous assets is used, with different sensing capabilities. Additionally, some or all of these assets may be operated by external organizations with an existing operations pipeline. To implement adaptive sampling in these systems, it could be prohibitively difficult to directly integrate into the existing planning software. Instead, they may expose a limited observation request interface that could be used. The federated planning paradigm, presented in this work, addresses this by using a multi-tiered planning approach to decouple the utility function from the individual asset operations pipelines. In this paradigm, the objective of the federated planner is to select a set of observations to request from each available asset such that the utility of the final collected observations are maximized, while minimizing the cost of acquiring those observations.


Simulation testing of federated planning for adaptive sampling and federated planning using a case study targeting the reduction in uncertainty in predictive hurricane modeling demonstrated a 4x increase in utility of traditional sampling methods [Branch et al. 2023].


Early prototyping of adaptive sampling federated planning algorithms started in 2020 under the POISE project. A simulation experiment of adaptive sampling applied to a case study of uncertainty reduction in predictive hurricane modeling was completed in 2022.


An example plan generated using federated planning to maximize a utility function.
An example plan generated using federated planning to maximize a utility function.
The federated planning problem is represented as a constraint optimization problem with inputs of a utility function, a set of assets to produce requests for, and a set of constraints, encoded as objective functions, related to those assets. These constraints utilize common specifications of the assets such as orbit path, max slew angle and rate, and autonomous drone endurance to determine potential observations and allow for the federated planners optimization algorithm to avoid selecting requests that are self-conflicting and infeasible for the individual asset planners to schedule. Due to the complexity of operating spacecraft and in-situ assets, it is generally expected that the federated planner will not be able to encode all operational constraints for all assets and thus will rely on the asset planners to account for all constraints and produce a final observation schedule. The POISE project developed a simulated annealing and greedy approach to the federated planning problem for adaptive sampling and compared them against traditional sampling methods.



Yuliya Marchetti
Andrew Branch
James Montgomery
Margaret Johnson
Longtao Wu
James Mason
Peyman Tavallali
Hui Su
Steve Chien


Office of the Chief Scientist and Chief Technologist Strategic Research and Technology Development Program