Abstract
As satellite constellations expand, the need for distributed, on-orbit intelligence becomes critical. However, traditional centralized training is hindered by massive data transmission costs and privacy concerns. Federated Learning (FL) offers a decentralized alternative, yet it struggles with the unique challenges of remote sensing: extreme class imbalance, non-independent and identically distributed (Non-IID) data, and foreground target scarcity. This paper presents SAFE (Self-Adjustment FEderated Learning), a framework designed to optimize collaborative perception through autonomous class rectification, feature alignment, and adaptive context enhancement.
Introduction
The paradigm of remote sensing is shifting toward "distributed space-based observation." While this provides unprecedented coverage, training robust AI models across a network of satellites is a logistical challenge. Beaming raw data to Earth for centralized training is slow, expensive, and poses security risks. Federated Learning—where satellites train models locally and share only weight updates—is a promising solution.
Despite its potential, standard FL algorithms like FedAvg often fail in the orbital environment. Satellites traversing different paths encounter vastly different distributions (Non-IID data); for instance, one node may observe only maritime environments while another sees urban centers. Furthermore, the objects of interest (foreground) in remote sensing images often occupy a negligible fraction of the total pixels compared to the background. The SAFE framework addresses these issues by introducing a self-adjusting mechanism that stabilizes training and enhances target sensitivity.
Methodology
The SAFE framework is built upon four core strategies designed to harmonize local learning with global convergence:
1. Class Rectification Optimization (CRO)
To combat global class imbalance without accessing private local data, the server maintains a small, balanced set of "Self-Examination Samples" (SES). By measuring how the global model reacts to these samples, the server can detect gradient "weakness" in underrepresented classes. It then autonomously adjusts loss weights to compensate for these gaps, ensuring that minority classes are not ignored during the aggregation process.
2. Feature Alignment Update (FAU)
Standard weight averaging can lead to "catastrophic forgetting" of local features. SAFE utilizes an Exponential Moving Average (EMA) update combined with Centered Kernel Alignment (CKA). This approach measures the feature-space similarity between the server and the client, allowing the model to retain beneficial local personalization while still absorbing global knowledge.
3. Dual-Factor Modulation Rheostat (DMR)
SAFE employs a curriculum-based training schedule. In the early stages of training, the DMR prioritizes global stability to establish a baseline. As the model matures, the "rheostat" shifts focus toward "hard" minority classes, ensuring the final model is robust across the entire data distribution rather than just the most frequent classes.
4. Adaptive Context Enhancement (ACE)
Remote sensing targets are often small and sparse. The ACE module is a plug-in component that utilizes trainable context vectors to generate a differentiable mask. This mask suppresses background noise (e.g., vast stretches of ocean or desert) and amplifies foreground features, significantly improving the detection of small-scale targets like ships, storage tanks, or specific infrastructure.
Experimental Setup and Results
The framework was rigorously tested on four major remote sensing benchmarks: PatternNet, NWPU-RESISC45, LoveDA, and WHDLD. To simulate real-world orbital conditions, the researchers used a Dirichlet distribution ($\\alpha = 0.5$) to create highly non-uniform data distributions across five virtual clients.
Performance Gains
SAFE consistently outperformed established baselines such as FedAvg and FedProx. On the NWPU-RESISC45 dataset, which is known for high intra-class diversity, SAFE achieved a 65.20% cloud class accuracy, representing a significant improvement over FedAvg’s 48.25%.
Robustness to Imbalance
The framework demonstrated remarkable stability even when the class imbalance ratio was pushed to 20:1. While baseline models saw rapid performance decay under these conditions, SAFE maintained high accuracy. Qualitative analysis via ACE heatmaps further confirmed that the model focuses more precisely on structural details—such as pier layouts and parking lot markings—than traditional FL methods, which are often distracted by environmental shadows or sea surfaces.
Discussion and Insights
The success of SAFE highlights two critical insights for the future of space-based AI:
- Gradients as Privacy-Preserving Proxies: The CRO module proves that the server does not need to see raw data to understand local struggles. Gradient analysis on a small validation set is sufficient to infer and rectify global imbalances.
- The Importance of Local Flavor: Forcing all satellites to converge to a single global average too quickly is counterproductive. By allowing for feature alignment rather than strict weight synchronization, SAFE preserves the specialized knowledge each satellite gains from its specific orbital path.
Limitations and Future Work
While SAFE represents a significant advancement, it relies on the availability of a small, balanced SES set at the server level, which may be difficult to acquire for highly specialized missions. Additionally, the framework currently assumes relatively homogeneous hardware across the network. Future research will focus on integrating SAFE with gradient compression techniques to accommodate narrow communication windows and exploring its performance across heterogeneous satellite hardware architectures.
Conclusion
The SAFE framework provides a robust solution for collaborative perception in distributed satellite networks. By addressing the inherent data imbalances and distribution shifts of remote sensing through autonomous rectification and context enhancement, SAFE enables high-performance AI training that is both data-efficient and privacy-preserving.
Summary updated