Abstract: With the rapid development of deep learning, the same set of backbone network architectures have often been adopted for multi-source remote sensing image classification and achieved impressive results. However, this raises severe security issues for remote sensing task as these models are known to be vulnerable to adversarial attacks. Conventional adversarial attack methods are able to fool single-band images classifier, but the adversarial perturbations of different band are not coupling, that results in its unavailability in multi-source remote sensing images attack. In this paper, we propose an adversarial attack method with consistent perturbation pattern across multi-source images, using sparse differential coevolution: firstly, generate some random vectors that contains the information of multi-source adversarial perturbations and generate the adversarial examples according to the vectors. Then generates new parameter vectors by crossover operator and mutation operator and generate the adversarial examples, too. Finally, evaluate the adversarial examples and choose the better vectors to survive. Repeat this progress and we can get the best multi-source adversarial examples. Experimental results demonstrate the feasibility of our strategy. Under one-pixel attack of our method, 61.38% of optical images and 38.93% of SAR images can be crafted to adversarial images.