2DDSRU-MobileNet: an end-to-end cloud-noise-robust lightweight convolution neural network

Abstract

To solve the problem of thin-cloud interference of remote sensing (RS) images under hardware-constrained environments, we present an end-to-end cloudnoise-robust lightweight convolution neural network model, 2DDSRU-MobileNet, based on MobileNetV3-small. We first propose a denoised module, which notably improves the robustness of lightweight convolution neural network under a thincloud environment. Then using the idea of a deep residual shrinkage network, we construct a two-dimensional deep shrinkage residual unit (2DDSRU). Then we introduce MobileNetV3-small as a background model to solve the problem of RS obstructed by thin cloud; this enables our method to obtain a higher efficiency compared with other methods. Moreover, deformable convolution supplants conventional dilated convolution, leading to agile capturing of contextual information in the thin-cloud image. Finally, the model's efficiency and accuracy are evaluated using the RSSCN-cloud and UCM-cloud datasets. The experimental results demonstrate the cloud-noise-robust performance of our model in the thin-cloud environment. Under a thin-cloud environment, the proposed 2DDSRU-MobileNet model surpasses classic models, such as AlexNet, VGG16, ResNet50, MobileNet series, ShuffleNet-v1, Efficientnet-b3, GhostNetV2, and MobileViT, in terms of runtime computational complexity, network parameters, and classification accuracy.