Dynamic convolutional model based on distribution-collaboration strategy for remote sensing scene classification

Most existing remote sensing image scene classification (RSISC) models rely mainly on cloud or terminal convolutional neural networks (CNN). However, with the advent of the era of high-resolution remote sensing image (HRRSI) big data, the accuracy and efficiency of such models on remote sensing multi-platforms are poor, which cannot satisfy the real-time and efficient image interpretation. This article proposes a novel RSISC model based on a distribution-collaboration strategy to address this problem. Firstly, the strategy allows most HRRSI to provide classification results in advance within a certain expected confidence range, reducing cloud transmission, reducing latency, and improving inference efficiency. Secondly, an adaptive enhanced attention mechanism based on the Lie Group feature covariance matrix is designed to aggregate the essential attribute feature (EAF) of HRRSI, which can effectively deal with HRRSI without increasing the computational complexity and delay of the model with the increase of HRRSI resolution. Finally, two types of dynamic convolution kernel (class consensus kernel and adaptive kernel) are proposed for scene classification. We chose to conduct extensive experiments on three challenging datasets and compared them with the state-of-the-art models, our model has certain advantages in classification accuracy, inference efficiency, computational performance, and number of parameters.