The Recognition of Remote Sensing Image based on Federated Generation Adversarial Network

Recently, the processing of Synthetic Aperture Radar(SAR) has been gradually transitioning from traditional electromagnetic simulation to intelligent detection and recognition. As an intelligent architecture, Federated Learning(FL) has attracted increasing attention for the processing of SAR images due to its capacity for data protection and cooperative gain. In this paper, a new FL scheme named Fed-CWGP is proposed, where Wasserstein Generation Adversarial Network with gradient penalty(WGAN-gp) is considered to expand samples locally for the issue of few-shot learning. Additionally, the Conditional Generation Adversarial Network(CGAN) is used to control the direction of sample generation, solving the problem of non-independent and non-identical distribution of data distribution between nodes. This proposed scheme allows local classifiers and cloud servers to exchange and collaboratively learn model parameters, which improves the performance of global and node classification models and reduces the transmission delay between models. To enhance the privacy of model uploading, differential privacy is introduced on the gradient of each node model. Simulation results state that the proposed Fed-CWGP outperforms the classical schemes.

I. Introduction

With the rapid development of satellite systems and remote sensing technology, the resolution of satellite remote sensing images has been greatly improved, and the task of remote sensing image classification has also changed. Remote sensing image classification is a key technology in remote sensing image application systems. Its primary purpose is to interpret and identify the type and distribution of surface objects according to the radiation information of ground object electromagnetic waves mapped to the feature information of remote sensing images. With the development of artificial intelligence, Machine Learning(ML) and Deep Learning(DL) have been used to classify remote-sensing images. For example, [1] is suitable for feature extraction of remote sensing images by using the strong capability of deep convolutional neural network for image feature recognition. However, these methods usually require a large amount of labeled training data or feature extraction from pre-trained CNN for fine-tuning, which is a very expensive and time-consuming behavior for remote sensing image data and requires the participation of experts. Therefore, most of us are faced with the problem that the sample size of labeled data is too small.