Source-linked AI summary
FedMAD: Modulation-Aware Directional Aggregation for Federated Learning in Remote Sensing Image Classification
Barış Büyüktaş, Begüm Demir
TL;DR
Heterogeneous client data can substantially degrade federated learning performance in remote sensing. FedMAD combines client-specific adaptation with aggregation based on local update alignment, and its reported results outperform state-of-the-art federated methods on BigEarthNet-S2 and EuroSAT.
Problem
Heterogeneous client data substantially degrades federated learning performance, and is common in remote sensing because image characteristics vary with acquisition conditions.
Source: §I · p. 1Method
FedMAD separates globally shared representation parameters from local client-specific modulation and batch-normalization parameters, and weights aggregation according to local modulation-update alignment.
Source: §I · pp. 1–2, A. Client-Specific Representation… · p. 2, B. Modulation-Aware Directional… · p. 2Results
FedMAD achieves 75.12% micro mAP on BigEarthNet-S2 and 67.14% overall accuracy on EuroSAT, outperforming FedAvg by approximately 14% on both datasets.
Source: D. Comparison with State-of-the-Art FL… · p. 4Takeaways & Limitations
Results demonstrate FedMAD's effectiveness compared with state-of-the-art federated learning algorithms for remote-sensing classification under heterogeneous data distributions.
Source: §I · p. 1
Abstract
from arXiv · showhide
Federated learning (FL) has recently attracted increasing attention in remote sensing (RS) since it enables collaborative model training across decentralized RS image archives without requiring direct access to local data. However, FL performance significantly degrades when the data distributions between clients are heterogeneous, which often occurs due to geographical differences, seasonal changes, and varying image acquisition and atmospheric conditions. To address this challenge, in this letter, we propose a novel personalized FL framework (denoted as FedMAD) for RS image classification problems. The proposed framework separates globally shared representation parameters from client-specific adaptation parameters to preserve client-specific features while maintaining globally transferable representations. This is achieved by integrating lightweight modulation modules and local batch normalization layers into the backbone network. Although globally shared parameters are collaboratively optimized between clients, client-specific parameters remain local to preserve domain-specific feature characteristics. In addition, FedMAD introduces a modulation-aware directional aggregation strategy that dynamically adjusts the importance of aggregation for each client according to the alignment of local modulation updates. This allows the global optimization process to suppress conflicting client updates originating from heterogeneous data distributions while enhancing the contribution of clients with consistent adaptation behaviors. The experimental results obtained on the BigEarthNet-S2 and EuroSAT datasets demonstrate the effectiveness of FedMAD compared to state-of-the-art FL algorithms under heterogeneous RS data distributions. The code of the proposed framework will be publicly available at https://git.tu-berlin.de/rsim/fedmad.
I. INTRODUCTION
Federated learning supports collaborative remote-sensing training without sharing local data, but heterogeneous client distributions can undermine performance. FedMAD combines client-specific feature adaptation with aggregation that accounts for alignment among local updates.
I. INTRODUCTION
- FL enables collaborative training across decentralized remote-sensing archives without direct access to local data.Centralizing local data may be infeasible because of privacy, legal, communication, or storage constraints.Source: §I · p. 1
- Heterogeneous client data substantially degrades FL performance, and is common in remote sensing because image characteristics vary with acquisition conditions.Conventional sample-count-based aggregation can produce unstable convergence and reduced generalization under non-IID data.Source: §I · p. 1
- Prior personalized FL methods preserve client-specific information through local normalization, personalized branches, or client-specific classifier heads.Most still treat client updates equally during global optimization, where conflicting updates may harm convergence and generalization.Source: §I · p. 1
- FedMAD combines globally shared parameters with client-specific adaptation, using lightweight modulation modules and local batch normalization to preserve domain-specific features.The framework is intended to maintain globally transferable representations while adapting to heterogeneous client data.Source: §I · pp. 1–2
- Its modulation-aware directional aggregation dynamically adjusts each client's contribution according to the alignment of local adaptation updates.This is intended to suppress conflicting updates during global optimization.Source: §I · pp. 1–2
II. PROPOSED FEDMAD FRAMEWORK
FedMAD addresses heterogeneous remote-sensing data distributions through two components: client-specific representation learning and modulation-aware directional aggregation.
II. PROPOSED FEDMAD FRAMEWORK
- FedMAD is designed to address limitations caused by heterogeneous remote-sensing data distributions during federated training.The framework has two main components: client-specific representation learning and modulation-aware directional aggregation.Source: §II · p. 2
- Client-specific representation learning is one of the framework's two main components.The other is modulation-aware directional aggregation.Source: §II · p. 2
- The framework presents the details of its two components in separate following subsections.Source: §II · p. 2
A. Client-Specific Representation Learning
FedMAD separates globally shared representation parameters from client-specific adaptation parameters, combining transferable representations with local feature adaptation. Lightweight modulation modules and local batch normalization support this separation while keeping client-specific parameters local.
A. Client-Specific Representation Learning
- Lightweight modulation modules apply client-specific feature adaptation while maintaining globally transferable image representations.The modulation parameters adjust feature activations according to each client's local data distribution.Source: A. Client-Specific Representation… · p. 2
- FedMAD keeps modulation and batch-normalization parameters local while collaboratively optimizing the remaining backbone parameters across clients.The modules are integrated before the backbone's batch-normalization layers.Source: A. Client-Specific Representation… · p. 2
- Client-specific parameters remain relatively small compared with the network, enabling personalized feature adaptation without noticeable computational overhead.Source: A. Client-Specific Representation… · p. 2
- Only globally shared parameters are sent for server aggregation, while client-specific parameters stay local so clients retain personalized feature representations.Both parameter groups are updated during local optimization.Source: A. Client-Specific Representation… · p. 2
B. Modulation-Aware Directional Aggregation
FedMAD addresses conflicting client updates under heterogeneous remote-sensing data by weighting aggregation according to how local modulation updates align. This is intended to suppress conflicting updates and emphasize clients with consistent adaptation behavior.
B. Modulation-Aware Directional Aggregation
- Heterogeneous data can make client optimization directions differ, producing conflicting updates during global optimization.FedMAD adjusts each client's aggregation importance according to alignment of local modulation updates.Source: B. Modulation-Aware Directional… · p. 2
- Local modulation updates characterize client-specific adaptation, and their directions are compared with a reference direction using cosine similarity.The reference direction aggregates local modulation directions across clients.Source: B. Modulation-Aware Directional… · p. 3
- FedMAD estimates client aggregation importance from alignment scores and uses those weights to update globally shared parameters.The aggregation weight also uses client training-set size, a numerical-stability constant, and a parameter controlling sensitivity to alignment scores.Source: B. Modulation-Aware Directional… · p. 3
- By suppressing conflicting updates and emphasizing clients with consistent adaptation behavior, the strategy is intended to improve optimization robustness under non-IID data.Source: B. Modulation-Aware Directional… · p. 3
A. Dataset Description
The evaluation uses BigEarthNet-S2 and EuroSAT, with client data arranged to represent heterogeneous remote-sensing settings. The datasets differ in their imagery and in how the client partitions are constructed.
A. Dataset Description
- BigEarthNet-S2 v2.0 and EuroSAT are the two evaluation datasets.Source: A. Dataset Description · p. 3
- BigEarthNet-S2 clients represented eight countries and used summer-season imagery; experiments used 10 m and 20 m bands, resizing only 20 m bands to 120×120 pixels by bicubic interpolation, yielding ten bands per image.The 60 m bands were excluded.Source: A. Dataset Description · p. 3
- EuroSAT uses RGB 64 × 64 images, with training samples distributed across clients by a Dirichlet distribution with α = 0.3.The data are split 80% for training and 20% for testing, and the resulting client label distributions are non-IID.Source: A. Dataset Description · p. 3
B. Design of Experiments
Experiments use ResNet50 and select the final configuration after testing modulation-layer counts and aggregation sensitivity. Training and evaluation settings are held to a specified communication schedule and dataset-specific metrics.
B. Design of Experiments
- The experiments use ResNet50 and vary modulation layers from 1 to 16 and aggregation parameter p from 0.5 to 3.0 before selecting a configuration for comparisons.Source: B. Design of Experiments · p. 3
- Training runs for 40 communication rounds, with two local epochs per round, using Adam at learning rate 0.001 and batch size 512.Source: B. Design of Experiments · p. 3
- Performance is evaluated by micro mean Average Precision on BigEarthNet-S2 and overall accuracy on EuroSAT, while computational complexity is assessed using FLOPs.Source: B. Design of Experiments · p. 3
C. Ablation and Sensitivity Study
The complete FedMAD configuration performs best among the reported ablation variants, while the component and sensitivity analyses show how modulation, aggregation, and hyperparameter choices relate to performance.
C. Ablation and Sensitivity Study
- 75.12% micro mAP on BigEarthNet-S2 and 67.14% on EuroSAT are achieved by the complete FedMAD framework.The authors report that client-specific adaptation and modulation-aware aggregation provide complementary benefits.Source: C. Ablation and Sensitivity Study · p. 4
- Adding modulation-aware aggregation to FedBN raises micro mAP from 72.42% to 73.43% on BigEarthNet-S2 and from 56.50% to 59.82% on EuroSAT.Private modulation modules with conventional aggregation yield 73.91% on BigEarthNet-S2 and 64.67% on EuroSAT.Source: C. Ablation and Sensitivity Study · p. 4
- Four modulation layers give the best BigEarthNet-S2 micro mAP of 75.12%, while p = 2.0 gives the best reported aggregation sensitivity setting.Performance is relatively stable for p values from 0.5 to 2.0; more modulation layers can reduce performance, and larger p can diminish other clients' contributions.Source: C. Ablation and Sensitivity Study · p. 4
D. Comparison with State-of-the-Art FL Methods
FedMAD outperforms the evaluated FL algorithms on BigEarthNet-S2 and EuroSAT while keeping computational complexity very close to FedAvg.
D. Comparison with State-of-the-Art FL Methods
- Table I evaluates private BN, private modulation, and modulation-aware aggregation; Table II examines modulation-layer count and alignment power p on BigEarthNet-S2.Source: Table I · p. 4, Table II · p. 4
- 75.12% micro mAP on BigEarthNet-S2 and 67.14% overall accuracy on EuroSAT; FedMAD achieves the highest performance on both datasets and beats FedAvg by approximately 14% on each.The comparison includes FedAvg, MOON, FedDC, FedAWA, FedGPS, and FedBN.Source: D. Comparison with State-of-the-Art FL… · p. 4, Table III · p. 5
- FedMAD’s modulation modules and directional aggregation are reported to improve robustness and generalization by preserving local adaptations and suppressing conflicting updates.The authors state that improvement over FedBN indicates that private normalization statistics alone are insufficient for modeling heterogeneous RS distributions.Source: D. Comparison with State-of-the-Art FL… · pp. 4–5
- FedMAD introduces negligible computational overhead, with complexity very close to FedAvg, FedBN, FedDC, and FedAWA.MOON and FedGPS have slightly higher FLOPs than the other compared methods.Source: D. Comparison with State-of-the-Art FL… · p. 5
IV. CONCLUSION
FedMAD separates globally shared representations from client-specific adaptations and uses modulation-aware directional aggregation for heterogeneous RS classification. The authors report gains over state-of-the-art FL methods with negligible computational overhead.
IV. CONCLUSION
- FedMAD separates globally shared representation parameters from client-specific adaptation parameters using lightweight modulation modules and local batch normalization.Source: §IV · p. 5
- A modulation-aware directional aggregation strategy adjusts client aggregation importance according to alignment of local adaptation behaviors.Source: §IV · p. 5
- Experiments on BigEarthNet-S2 and EuroSAT show that FedMAD consistently outperforms state-of-the-art FL algorithms while adding negligible computational overhead.Source: §IV · p. 5
- The modulation modules and aggregation strategy are independent of network architecture and can be integrated into different deep-learning backbones and FL systems.Source: §IV · p. 5
- Future work will explore adaptive mechanisms for determining the degree of personalization and knowledge sharing during FL.Source: §IV · p. 5