Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Federated Learning Across Decentralized and Unshared Archives for Remote Sensing Image Classification

Barış Büyüktaş, Gencer Sumbul, Begüm Demir

TL;DR

This paper addresses federated learning for remote sensing image classification across decentralized and unshared archives by conducting the first RS-focused comparative study of state-of-the-art FL algorithms. It combines a systematic CV/ML FL survey with a RS-centric theoretical comparison across local training complexity, aggregation complexity, learning efficiency, communication cost, and scalability, followed by extensive RS experiments on BigEarthNet-S2 using three decentralization scenarios. Key findings show that local-training-focused methods (e.g., MOON, FedDC) offer robustness to non-IID data, while aggregation-focused approaches (e.g., FedNova, FedBN, pFedLA) deliver competitive performance with different cost-profile trade-offs; MOON provides strong early performance but at higher computation, whereas FedBN/FedDC balance robustness and cost. The work delivers practical guidelines for RS FL deployment, highlights the potential for privacy-preserving and multi-modal RS data integration, and provides public code to enable broader RS FL research and applications.

Abstract

Federated learning (FL) enables the collaboration of multiple deep learning models to learn from decentralized data archives (i.e., clients) without accessing data on clients. Although FL offers ample opportunities in knowledge discovery from distributed image archives, it is seldom considered in remote sensing (RS). In this paper, as a first time in RS, we present a comparative study of state-of-the-art FL algorithms for RS image classification problems. To this end, we initially provide a systematic review of the FL algorithms presented in the computer vision and machine learning communities. Then, we select several state-of-the-art FL algorithms based on their effectiveness with respect to training data heterogeneity across clients (known as non-IID data). After presenting an extensive overview of the selected algorithms, a theoretical comparison of the algorithms is conducted based on their: 1) local training complexity; 2) aggregation complexity; 3) learning efficiency; 4) communication cost; and 5) scalability in terms of number of clients. After the theoretical comparison, experimental analyses are presented to compare them under different decentralization scenarios. For the experimental analyses, we focus our attention on multi-label image classification problems in RS. Based on our comprehensive analyses, we finally derive a guideline for selecting suitable FL algorithms in RS. The code of this work is publicly available at https://git.tu-berlin.de/rsim/FL-RS.

Federated Learning Across Decentralized and Unshared Archives for Remote Sensing Image Classification

TL;DR

This paper addresses federated learning for remote sensing image classification across decentralized and unshared archives by conducting the first RS-focused comparative study of state-of-the-art FL algorithms. It combines a systematic CV/ML FL survey with a RS-centric theoretical comparison across local training complexity, aggregation complexity, learning efficiency, communication cost, and scalability, followed by extensive RS experiments on BigEarthNet-S2 using three decentralization scenarios. Key findings show that local-training-focused methods (e.g., MOON, FedDC) offer robustness to non-IID data, while aggregation-focused approaches (e.g., FedNova, FedBN, pFedLA) deliver competitive performance with different cost-profile trade-offs; MOON provides strong early performance but at higher computation, whereas FedBN/FedDC balance robustness and cost. The work delivers practical guidelines for RS FL deployment, highlights the potential for privacy-preserving and multi-modal RS data integration, and provides public code to enable broader RS FL research and applications.

Abstract

Federated learning (FL) enables the collaboration of multiple deep learning models to learn from decentralized data archives (i.e., clients) without accessing data on clients. Although FL offers ample opportunities in knowledge discovery from distributed image archives, it is seldom considered in remote sensing (RS). In this paper, as a first time in RS, we present a comparative study of state-of-the-art FL algorithms for RS image classification problems. To this end, we initially provide a systematic review of the FL algorithms presented in the computer vision and machine learning communities. Then, we select several state-of-the-art FL algorithms based on their effectiveness with respect to training data heterogeneity across clients (known as non-IID data). After presenting an extensive overview of the selected algorithms, a theoretical comparison of the algorithms is conducted based on their: 1) local training complexity; 2) aggregation complexity; 3) learning efficiency; 4) communication cost; and 5) scalability in terms of number of clients. After the theoretical comparison, experimental analyses are presented to compare them under different decentralization scenarios. For the experimental analyses, we focus our attention on multi-label image classification problems in RS. Based on our comprehensive analyses, we finally derive a guideline for selecting suitable FL algorithms in RS. The code of this work is publicly available at https://git.tu-berlin.de/rsim/FL-RS.
Paper Structure (27 sections, 17 equations, 3 figures, 6 tables, 2 algorithms)

This paper contains 27 sections, 17 equations, 3 figures, 6 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Recent Advances in Federated Learning Addressing Training Data Heterogeneity Across Clients
  3. Recent Advances in Federated Learning for Computer Vision and Machine Learning
  4. Recent Advances in Federated Learning for Remote Sensing
  5. An Overview of the Selected Federated Learning Algorithms
  6. The Local Training Focused Algorithms
  7. FedProx li2020federated
  8. MOON li2021model
  9. SCAFFOLD karimireddy2020scaffold
  10. FedDC gao2022feddc
  11. The Model Aggregation Focused Algorithms
  12. FedNova wang2021novel
  13. FedBN li2021fedbn
  14. pFedLA ma2022layer
  15. A Theoretical Comparison of the Selected Federated Learning Algorithms
  16. ...and 12 more sections

Figures (3)

  • Figure 1: An illustration of federated learning applied with the iterative model averaging strategy to learn a deep neural network (DNN) when there is no access to training data on decentralized RS image archives (i.e., clients). There are four main steps of iterative model averaging: ① each client updates its DNN with the corresponding local data, leading to one set $w_k$ of local DNN parameters for the $k$th client; ② each client sends its local DNN parameters to the central server; ③ the central server aggregates the parameters of all local models, aiming to obtain the set $w^*$ of optimum parameters of the global DNN; and ④ the central server sends the parameters of the global DNN to each client. These steps are repeated several times until the convergence of the global DNN.
  • Figure 2: $F_1$ score versus communication round obtained by the FedProx algorithm when the different numbers $E$ of local epochs are considered under the DS1.
  • Figure 3: $F_1$ score versus communication round obtained by the investigated FL algorithms under the DS1.