LABANet: A Lightweight Asymmetrical Bottleneck and Attention-Based Network for Cloud Detection

Recently, onboard satellite cloud detection has attracted extensive interest due to its ability to reduce invalid data and save satellite storage and downlink resources. As the demand for accuracy has increased, deep neural networks with encoder–decoder architectures have been widely used for cloud detection. Most networks mainly focus on extracting contextual features and fusing more information derived from multilevel features to achieve high accuracy. Convolutional layers with contextual feature extraction abilities in networks, however, always contain many computations and parameters, making them difficult to deploy in processors with limited computing and memory resources. Additionally, fusing features acquired from different levels easily introduces redundant noise and decreases accuracy. To address these issues, a novel lightweight asymmetrical bottleneck and attention-based network (LABANet) is proposed. In the encoder, a multiscale context-based asymmetrical bottleneck (MCAB) is designed to extract more contextual features in an efficient manner. The number of computations and parameters required by the bottleneck can be effectively reduced via asymmetric channel compression and grouping for different pointwise convolutions. Multiscale contextual features can be extracted by dilated depthwise asymmetric convolution. In the decoder, a spatial-channel cascade attention (SCCA) module is proposed to fuse multilevel features with less redundant noise by highlighting spatial and channel correlation information. Experimental results obtained on public datasets demonstrate that the proposed LABANet can exceed 95% overall accuracy (OA) with a computational cost of 1.20 giga-floating-point operations (GFLOPs) and a parameter count of
$1.2\times 10^{{6}}$
; thus, it achieves competitive accuracy with fewer computations and parameters than other baseline cloud detection networks.