AKA-MobileNet: A Cloud-Noise-Robust lightweight Convolution Neural Network

In order to solve the problem of thin cloud interference of remote sensing images under hardware-constrained environments, this paper presents a new cloud-noise-robust lightweight convolution neural network model, AKA-MobileNet, based on MobileNetV3-small. The paper firstly proposes a new extract feature module. The new extract feature unit can notably improve the robustness of lightweight convolution neural network under thin cloud environment. The idea of AKConv (convolutional kernel with arbitrary sampled shapes and arbitrary number of parameters) enlightens us to construct a new unit named AKANet (The multi-scale feature composed of features extracted by AKConv). Then MobileNetV3-small is introduced as a background model to solve the problem of RS (remote sensing) to be obstructed by thin cloud. It makes our method to obtain higher efficiency compared other method. Finally, the model's efficiency and accuracy were evaluated using the RSSCN-cloud datasets. The experimental results demonstrate the cloud-noise-robust performance of our model in the thin cloud environment. Under thin cloud environment, the proposed AKC-MobileNet model surpasses classic model like AlexNet, VGG16, ResNet50, MobileNet series, ShuffleNet-v1 and Efficientnet-b3 in terms of runtime computational complexity, network’s parameters, and classification accuracy.