.png){kind=icon}
.png){kind=icon}
Return
Federated Coupled Contrastive Learning on Remote Sensing Image Classification
.png){kind=icon}
PRE
AI
DOI：10.1109/TGRS.2026.3654941.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Classification for remote sensing (RS) images plays a crucial role in various fields such as disaster assessment and urban planning. Nevertheless, due to resource limitations and privacy issues, directly transferring data to the ground for high-performance centralized training is often impractical. Federated learning (FL) has emerged as a revolutionary distributed learning paradigm, which enables collaborative training of a model without data sharing. However, since the RS data is distributed across numerous participants, FL suffers from an inevitable challenge of data heterogeneity for RS image classification. As the local data is more unevenly distributed, this challenge becomes increasingly pronounced, leading to inconsistent local update directions. In this article, we propose RS-CCL, a novel federated coupled contrastive learning rule on RS image classification, which expands the feature space, thereby allowing models to better learn the intrinsic structure of the complex RS data effectively. Our RS-CCL mathematically formulates cross-entropy and feature similarity into a unified framework. It is then solved by coupled gradient computation in a consistent optimization direction, overcoming dimensional collapse. In this way, an enriched feature space is created, thereby facilitating the training of more capable models. Within this framework, we further introduce a dynamic fusion mask matrix, which inhibits the model from excessively drawing on interclass similarity. Empirically, we show that our RS-CCL is in line with robustness theory that a larger feature space brings more robust performance. Extensive experiments on several benchmarks validate the effectiveness and generalization ability of our method. For example, with a high data heterogeneity of $\alpha =0.05$ , our RS-CCL outperforms other baselines by up to 8.5% on the UC Merced Land Use (UC Merced) dataset with MobileViT-S. The code is available at https://github.com/cxy-xd/RS-CCL
Keywords:
Federated learning (FL)
remote sensing (RS)
supervised contrastive learning (SCL)
Journal
IF:
8.6
Papers:
2.1W
Citations:
10.7W