PiCSRL: Physics-Informed Contextual Spectral Reinforcement Learning
Summary: arXiv:2603.26816v1 Announce Type: cross
Abstract
High-dimensional low-sample-size (HDLSS) datasets pose significant challenges to the development of reliable environmental models, primarily due to the scarcity of labeled data. Reinforcement learning (RL)-based adaptive sensing methods have the potential to learn optimal sampling policies; however, their effectiveness is severely limited in HDLSS contexts. In this study, we introduce PiCSRL (Physics-Informed Contextual Spectral Reinforcement Learning), which leverages domain knowledge to design embeddings that are directly parsed into the RL state representation, thereby enhancing adaptive sensing capabilities.
Methodology
Our approach includes the development of an uncertainty-aware belief model that integrates physics-informed features for improved predictive accuracy. As a representative application, we evaluated PiCSRL in the context of adaptive sampling for cyanobacterial gene concentration using NASA’s PACE hyperspectral imagery over Lake Erie.
Results
PiCSRL demonstrated exceptional performance in optimal station selection, achieving a Root Mean Square Error (RMSE) of 0.153 and a 98.4% bloom detection rate. This is a significant improvement compared to the baseline methods, including:
- Random Selection: RMSE = 0.296
- Upper Confidence Bound (UCB): RMSE = 0.178
Our ablation experiments further revealed that the integration of physics-informed features results in enhanced test generalization, achieving an R² value of 0.52, which is an improvement of 0.11 over raw spectral bands in a semi-supervised learning framework.
Scalability
In addition to its effectiveness, PiCSRL has shown robust scalability in larger networks. Our scalability tests indicate that PiCSRL can efficiently handle networks comprising 50 stations and more than 2 million combinations, showcasing significant improvements over baseline models (p = 0.002).
Conclusion
We propose PiCSRL as a sample-efficient adaptive sensing method that can be applied across various Earth observation domains, ultimately leading to improved observation-to-target mapping. The integration of physics-informed features not only enhances the efficacy of adaptive sensing but also sets a new benchmark for future research in this field.
As environmental data becomes increasingly important in addressing global challenges, methods like PiCSRL could play a critical role in advancing our understanding and management of ecological systems.