SAFE: Self-Adjustment Federated Learning Framework for Remote Sensing Collaborative Perception

TL;DR

SAFE tackles data privacy, communication overhead, and distribution shift in distributed remote sensing by introducing four mechanisms—CRO, FAU, DMR, and ACE—that coordinate local and global learning. CRO estimates global class proportions via gradients and reweights losses without sharing data, while FAU preserves local representations through EMA-based updates guided by cross-domain feature alignment. DMR provides a curriculum-like balance between global convergence and local personalization, and ACE enhances foreground representation with a lightweight, end-to-end foreground augmentation. Across classification and segmentation benchmarks under non-IID and imbalanced conditions, SAFE delivers substantial gains over standard FL baselines, demonstrating robust adaptability and practical utility for on-orbit collaborative sensing.

Abstract

The rapid increase in remote sensing satellites has led to the emergence of distributed space-based observation systems. However, existing distributed remote sensing models often rely on centralized training, resulting in data leakage, communication overhead, and reduced accuracy due to data distribution discrepancies across platforms. To address these challenges, we propose the \textit{Self-Adjustment FEderated Learning} (SAFE) framework, which innovatively leverages federated learning to enhance collaborative sensing in remote sensing scenarios. SAFE introduces four key strategies: (1) \textit{Class Rectification Optimization}, which autonomously addresses class imbalance under unknown local and global distributions. (2) \textit{Feature Alignment Update}, which mitigates Non-IID data issues via locally controlled EMA updates. (3) \textit{Dual-Factor Modulation Rheostat}, which dynamically balances optimization effects during training. (4) \textit{Adaptive Context Enhancement}, which is designed to improve model performance by dynamically refining foreground regions, ensuring computational efficiency with accuracy improvement across distributed satellites. Experiments on real-world image classification and object segmentation datasets validate the effectiveness and reliability of the SAFE framework in complex remote sensing scenarios.

Figures (9)

• Figure 1: In the federated learning framework, each satellite terminal independently trains its model using locally collected datasets. During each communication cycle, the locally trained parameters are uploaded to a central server, which aggregates them to update the global model. The updated global parameters are then sent back to the terminals for further training.
• Figure 2: The proposed SAFE framework adheres to the workflow of the federated learning paradigm. A small set of Self-Examination Samples is utilized at the central node to compute Gradient Proportion and Feature Alignment. The former addresses class bias correction during local training on terminals, while the latter guides the extent of EMA updates for local models. A Dual-Factor Modulation mechanism regulates the entire training process. Additionally, the client embeddings stored on each terminal are used to guide the Adaptive Context Enhancement mechanism for foreground enhancement.
• Figure 3: The Exponential Moving Average (EMA) update ratios of model parameters between clients and the cloud under the Dual-Factor Modulation Rheostat (DMR) mechanism.
• Figure 4: The proposed Adaptive Context Enhancement (ACE) module is designed to enhance the foreground regions within the feature extraction network in an end-to-end manner. Each terminal optimizes a dedicated client embedding, while cross-attention and differentiable gumbel sampling are employed to identify foreground regions and enhance them during feature forward propagation.
• Figure 5: The distribution of classes across clients in the PatternNet, NWPU-RESISC45, LoveDA and WHDLD datasets.