Geometric Knowledge-Assisted Federated Dual Knowledge Distillation Approach Towards Remote Sensing Satellite Imagery

Luyao Zou; Fei Pan; Jueying Li; Yan Kyaw Tun; Apurba Adhikary; Zhu Han; Hayoung Oh

cs.CVMar 8, 2026

Geometric Knowledge-Assisted Federated Dual Knowledge Distillation Approach Towards Remote Sensing Satellite Imagery

Authors: Luyao Zou, Fei Pan, Jueying Li, Yan Kyaw Tun, Apurba Adhikary, Zhu Han, Hayoung Oh

Organizations: College of Computing and Informatics, Sungkyunkwan University, South Korea · Department of Computer Science and Engineering, University of Michigan, United States · Konkuk Aerospace Design-Airworthiness Institute, Konkuk University, Seoul, 05029, South Korea · Department of Electronic Systems, Aalborg University, 2450 København SV , Denmark ·

Abstract

Federated learning (FL) has recently become a promising solution for analyzing remote sensing satellite imagery (RSSI). However, the large scale and inherent data heterogeneity of images collected from multiple satellites, where the local data distribution of each satellite differs from the global one, present significant challenges to effective model training. To address this issue, we propose a Geometric Knowledge-Guided Federated Dual Knowledge Distillation (GK-FedDKD) framework for RSSI analysis. In our approach, each local client first distills a teacher encoder (TE) from multiple student encoders (SEs) trained with unlabeled augmented data. The TE is then connected with a shared classifier to form a teacher network (TN) that supervises the training of a new student network (SN). The intermediate representations of the TN are used to compute local covariance matrices, which are aggregated at the server to generate global geometric knowledge (GGK). This GGK is subsequently employed for local embedding augmentation to further guide SN training. We also design a novel loss function and a multi-prototype generation pipeline to stabilize the training process. Evaluation over multiple datasets showcases that the proposed GK-FedDKD approach is superior to the considered state-of-the-art baselines, e.g., the proposed approach with the Swin-T backbone surpasses previous SOTA approaches by an average 68.89% on the EuroSAT dataset.

Figures & tables

Fig. 1 : The discrepancy between local & global distribution for satellite imagery. Each satellite may have different categories and different amounts of data. Images are derived from the Satellite Image Classification (SIC) Dataset [ 5 ] .

Fig. 2 : Architecture of the proposed GK-FedDKD method, which is formed by multiple clients (i.e., satellites) and a server. This architecture showcases the specific implementation of Fig. 3 .

Fig. 3 : The overall framework of the proposed GK-FedDKD method includes: client & server parts. The client part contains: A KD with unlabeled augmented data for TE generation, B KD with labeled data and local covariance matrices calculation, C GVs-guided local embedding augmentation (LEA) and local learning regulation, D a linear layer-based module, and E multi-prototype (MP) generation strategy. The server part plays the role of 1) global covariance matrices generation, 2) model aggregation, and 3) MP aggregation.

Panel of Figure 4 — Fig. 4 : Per-class data sample distribution across ten clients.

Panel of Figure 5 — Fig. 5 : Data augmentation over the SIC dataset with rotation, Gaussian noise, flip, brighter, darker, saturation, and salt and pepper, respectively.

Panel of Figure 6 — Fig. 6 : Convergence comparison among FedExP, FedAS, FedPer and the proposed GK-FedDKD method.

BB	Method	EuroSAT			SAT4			SAT6
		Accuracy	AA	MFS	Accuracy	AA	MFS	Accuracy	AA	MFS
Swin-T	FedExP [ 23 ]	$0.6001$	$0.6046$	$0.5877$	$0.5656$	$0.5896$	$0.5206$	$0.8875$	$0.8652$	$0.7301$
	MOON [ 54 ]	$0.7179$	$0.7187$	$0.7016$	$0.3506$	$0.3194$	$0.1298$	$0.7335$	$0.7084$	$0.5310$
	FedAU [ 24 ]	$0.6713$	$0.6785$	$0.6432$	$0.7786$	$0.7923$	$0.7731$	$\color[rgb]{0,0,1}{0.9240}$	$\color[rgb]{0,0,1}{0.8997}$	$\color[rgb]{0,0,1}{0.7913}$
	FedProto [ 13 ]	$\color[rgb]{0,0,1}{0.8286}$	$\color[rgb]{0,0,1}{0.8281}$	$\color[rgb]{0,0,1}{0.8203}$	$\color[rgb]{0,0,1}{0.8246}$	$\color[rgb]{0,0,1}{0.8375}$	$\color[rgb]{0,0,1}{0.8130}$	$0.8740$	$0.8678$	$0.7495$
	FedProx [ 55 ]	$0.2156$	$0.1348$	$0.1294$	$0.4992$	$0.5257$	$0.3955$	$0.4075$	$0.3247$	$0.2067$

TABLE I : Performance comparison among the proposed GK-FedDKD and other baselines (BB: Backbone; AA: Average Accuracy; MAS: Macro-F1 Score). The highest values are shown in bold, the second best is marked in blue, and the difference between the best and the second best is depicted in red.

Dataset	Metrics	Method
		FedExP [ 23 ]	MOON [ 54 ]	FedProto [ 13 ]	FedAU [ 24 ]	Ours	$\bigtriangleup$
EuroSAT	Error Rate	$0.3999$	$0.2821$	$\color[rgb]{0,0,1}{0.1714}$	$0.3287$	0.0997	$-0.0717$
	MAE	$1.7178$	$1.1016$	$\color[rgb]{0,0,1}{0.6726}$	$1.3353$	0.3710	$-0.3016$
SIC	Error Rate	$\color[rgb]{0,0,1}{0.0966}$	$0.1572$	$0.1452$	$0.2406$	0.0768	$-0.0198$
	MAE	$\color[rgb]{0,0,1}{0.1044}$	$0.1710$	$0.1518$	$0.2567$	0.0840	$-0.0204$
SAT4	Error Rate	$0.4344$	$0.6494$	$\color[rgb]{0,0,1}{0.1754}$	$0.2214$	0.0484	$-0.127$

TABLE II : Error rate and mean absolute error (MAE) achieved by the proposed method and other baselines. Bold numbers denote the best values, and blue numbers represent the second best. Notably, the lower the value in this Table, the better the result.

Dataset	MPG	LL_M	GIA	TE Generation	Swin-T
					Accuracy	$\bigtriangleup$
EuroSAT					$0.8977$	$-0.0026$
					$0.8958$	$-0.0045$
					$0.8974$	$-0.0029$
					$0.8961$	$-0.0042$
					$0.9003$	$-$

TABLE III : Importance of critical components for the proposed method (MPG: multi-prototype generation, LL_M: linear layer-based module, GIA: global information alignment).

LR	Accuracy
	FedAS	FedPer	Ours	$\bigtriangleup$
0.01	$0.8445$	$\color[rgb]{0,0,1}{0.8490}$	0.9650	$\color[rgb]{1,0,0}{+0.116}$
0.001	$0.7415$	$\color[rgb]{0,0,1}{0.7915}$	0.9485	$\color[rgb]{1,0,0}{+0.157}$
0.0001	$0.5155$	$\color[rgb]{0,0,1}{0.6860}$	0.9175	$\color[rgb]{1,0,0}{+0.2315}$

TABLE IV : Accuracy obtained by FedAS, FedPer and the proposed GK-FedDKD approach when considering various learning rates for the SAT6 dataset. The best value is highlighted in bold, the second best is shown in blue, and the enhancement is offered in red.

Fig. 7 : The convergence curve was obtained by using the proposed GK-FedDKD approach over SAT6 dataset under the consideration of $LR=0.01$ and $LR=0.001$ .

Method	# $10$	# $20$	# $30$	# $50$
FedExP [ 23 ]	$0.8875$	$0.8387$	$\color[rgb]{0,0,1}{0.9045}$	$\color[rgb]{0,0,1}{0.9270}$
MOON [ 54 ]	$0.7335$	$0.7521$	$0.7536$	$0.5975$
FedAU [ 24 ]	$\color[rgb]{0,0,1}{0.9240}$	$0.8280$	$0.6433$	$0.5782$
FedProto [ 13 ]	$0.8740$	$\color[rgb]{0,0,1}{0.8857}$	$0.8567$	$0.7560$
FedProx [ 55 ]	$0.4075$	$0.5726$	$0.7034$	$0.6870$
FedAS [ 25 ]	$0.7415$	$0.7746$	$0.5960$	$0.5517$

TABLE V : Accuracy achieved by the proposed method and the baselines under the consideration of various numbers of clients (i.e.,

\{10,20,30,50\}

), where SAT6 dataset is adopted. The optimal values are shown in bold, and the second-best is marked in blue. The improvements between them are provided in red.

Method	$Dir=0.5$			$Dir=50$
	Accuracy	AA	MFS	Accuracy	AA	MFS
FedExP [ 23 ]	$0.8665$	$0.8397$	$\color[rgb]{0,0,1}{0.7564}$	$\color[rgb]{0,0,1}{0.9325}$	$\color[rgb]{0,0,1}{0.9346}$	$\color[rgb]{0,0,1}{0.8261}$
MOON [ 54 ]	$0.7595$	$0.7299$	$0.4856$	$0.8945$	$0.8940$	$0.7222$
FedAU [ 24 ]	$0.88$	$0.8419$	$0.6836$	$0.9320$	$0.9321$	$0.7577$
FedProto [ 13 ]	$\color[rgb]{0,0,1}{0.8855}$	$\color[rgb]{0,0,1}{0.8610}$	$0.7417$	$0.9265$	$0.9263$	$0.8252$
FedProx [ 55 ]	$0.5535$	$0.5591$	$0.3372$	$0.3490$	$0.3491$	$0.1745$

TABLE VI : Performance comparison among the proposed GK-FedDKD and other baselines (AA: Average Accuracy; MAS: Macro-F1 Score) over SAT6 dataset. The largest values are depicted in bold, the second-best is highlighted in blue, and the improvement between the best and the second-best is shown in red.

Panel of Figure 14 — Fig. 8 : Local embedding visualization for the test part of SAT6 dataset. The corresponding local embedding is generated by using the global model trained by the proposed method in round $30$ .

Panel of Figure 15 — Fig. 9 : Confusion matrix achieved by leveraging the proposed GK-FedDKD approach.

Federated Knowledge Distillation

Federated knowledge distillation transfers predictive behavior among data-isolated clients to train useful models without sharing raw data.

See all 36 papers

FedMAD: Modulation-Aware Directional Aggregation for Federated Learning in Remote Sensing Image Classification
Sep 30, 2026Barış Büyüktaş, Begüm DemirRemote Sensing Image Understanding Federated Learning Aggregation
Faculty of Electrical Engineering and Computer Science, Technische Universität Berlin, 10587 Berlin, Germany · BIFOLD - Berlin Institute for the Foundations of Learning and Data, 10587 Berlin, Germany
The Impact of Federated Learning on Distributed Remote Sensing Archives
Apr 13, 2026Anand Umashankar, Karam Tomotaki-Dawoud, Nicolai SchneiderFederated Learning Non-IID Federated Learning
Technische Universität Berlin Berlin, Germany
FedEU: Evidential Uncertainty-Driven Federated Fine-Tuning of Vision Foundation Models for Remote Sensing Image Segmentation
Mar 8, 2026Xiaokang Zhang, Xuran Xiong, Jianzhong Huang +1Federated Fine-Tuning Remote Sensing Image Segmentation
School of Artificial Intelligence, Wuhan University, Wuhan 430072, China · School of Information Science and Engineering, Wuhan University of Science and Technology, Wuhan 430081, China · School of Computer Science, Wuhan University, Wuhan 430072, China

Sep 30, 2026J/K move · Enter open · S save

FedMAD: Modulation-Aware Directional Aggregation for Federated Learning in Remote Sensing Image Classification

Barış Büyüktaş, Begüm Demir

Faculty of Electrical Engineering and Computer Science, Technische Universität Berlin, 10587 Berlin, Germany · BIFOLD - Berlin Institute for the Foundations of Learning and Data, 10587 Berlin, Germany

Remote Sensing Image Understanding Federated Learning Aggregation Non-IID Federated Learning Federated Learning Heterogeneous Federated Learning

Open paper AskPDF arXiv

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.

Abstract

Federated Knowledge Distillation

Explore similar work

FedMAD: Modulation-Aware Directional Aggregation for Federated Learning in Remote Sensing Image Classification

The Impact of Federated Learning on Distributed Remote Sensing Archives

FedEU: Evidential Uncertainty-Driven Federated Fine-Tuning of Vision Foundation Models for Remote Sensing Image Segmentation