Advanced federated learning methods for the analysis of remote sensing images across decentralized and unshared archives
Advances in remote sensing (RS) technologies have resulted in the rapid increase in the volume of RS images distributed across decentralized archives (i.e., clients). While deep learning (DL) has achieved remarkable success in extracting knowledge from large-scale RS data, most existing DL methods rely on centralized access to archives. However, RS images might be distributed across different clients, and these images may remain unshared due to commercial interests, legal regulations, and privacy concerns, making conventional DL methods infeasible in many RS applications. To overcome this challenge, federated learning (FL) can be used, which enables clients to collaboratively train models on decentralized archives while preserving data locality and privacy through the aggregation of locally computed updates at a central server. Despite its potential, existing FL methods face several limitations when applied to RS image classification, particularly regarding: i) challenges in mitigating the effects of training data heterogeneity; ii) the limited ability of FL to exploit complementary information from multi-modal data distributed across clients; and iii) the communication overhead arising from the frequent exchange of large-scale model updates across clients and a central server. This thesis addresses these challenges by introducing advanced FL strategies designed for RS image classification. First, we present the first comparative study on advanced FL algorithms in the context of RS image classification. Building on a systematic review, we select several state-of-the-art FL algorithms and assess their ability to address training data heterogeneity. We provide both theoretical and experimental comparisons under diverse decentralization scenarios, evaluating the algorithms with respect to local training complexity, aggregation complexity, learning efficiency, communication cost, and scalability. The findings demonstrate limitations of conventional FL algorithms when applied to heterogeneous RS archives. Based on these analyses, we provide a guideline for selecting suitable FL algorithms for RS image classification. Second, we propose a novel multi-modal FL framework that learns DL model parameters from decentralized and unshared RS image archives associated with different modalities (e.g., Sentinel-1 and Sentinel-2). The framework integrates three modules: 1) multi-modal fusion module to preserve complementary information across clients; 2) feature whitening module to align heterogeneous feature distributions and mitigate training data heterogeneity; and 3) mutual information maximization module to enforce cross-modal consistency without requiring direct access to paired multi-modal samples. The joint use of all modules enables the framework to fully exploit complementary information from decentralized multi-modal data, while effectively mitigating heterogeneity across clients. Third, we introduce FedX, an explanation-guided pruning strategy that leverages backpropagation-based interpretability methods to assess the relevance of model parameters and remove less informative components from the global model. By identifying and retaining only the most task-relevant parameters, FedX enables the transmission of sparse global models that substantially reduce communication overhead while maintaining strong generalization performance. Beyond improving efficiency, FedX demonstrates superior robustness under high sparsity levels compared to state-of-the-art pruning approaches, highlighting its effectiveness for communication-constrained FL scenarios in RS. These contributions of this thesis provide a foundation for advancing FL methods for RS by enabling scalable, communication-efficient, and privacy-preserving learning from decentralized and unshared RS image archives. The code for all contributions of this thesis is made publicly available to support reproducibility and further research.
