A Two-Stage Federated Learning Framework for Class Imbalance in Aerial Scene Classification

Centralized aerial imagery analysis techniques face two challenges. The first one is the data silos problem, where data is located at different organizations separately. The second challenge is the class imbalance in the overall distribution of aerial scene data, due to the various collecting procedures across organizations. Federated learning (FL) is a method that allows multiple organizations to learn collaboratively from their local data without sharing. This preserves users' privacy and tackles the data silos problem. However, traditional FL methods assume that the datasets are globally balanced, which is not realistic for aerial imagery applications. In this paper, we propose a Two-Stage FL framework (TS-FL), which mitigate the effect of the class imbalanced problem in aerial scene classification under FL. In particular, the framework introduces a feature representation method by combing supervised contrastive learning with knowledge distillation to enhance the model's feature representation ability and minimize the client drift. Experiments on two public aerial datasets demonstrate that the proposed method outperforms other FL methods and possesses good generalization ability.