High-resolution remote sensing images with rich land surface structure can provide data support for accurately understanding more detailed change information of land cover and land use (LCLU) at different times. In this study, we present a novel scene change understanding framework for remote sensing which includes scene classification and change detection. To enhance the feature representation of images in scene classification, a robust label semantic relation learning (LSRL) network based on EfficientNet is presented for scene classification. It consists of a semantic relation learning module based on graph convolutional networks and a joint expression learning framework based on similarity. Since the bi-temporal remote sensing image pairs include spectral information in both temporal and spatial dimensions, land cover and land use change monitoring can be improved by using the relationship between different spatial and temporal locations. Therefore, a change detection method based on swin transformer blocks (STB-CD) is presented to obtain contextual relationships between targets. The experimental results on the LEVIR-CD, NWPU-RESISC45, and AID datasets demonstrate the superiority of LSRL and STB-CD over other state-of-the-art methods.

Detection of streak-like space targets in optical star images is of great importance for space situational awareness and on-orbit asset security. However, when the signal-to-noise ratio of streaks is low, they are obscured by background stars and stray light, causing traditional direct segmentation methods to miss the target. To enhance the ability of faint streak detection, anisotropic Gaussian kernels with grayscale-similarity adaptive weighting are firstly used to extract the local direction information and suppress the influence of stars. Secondly, a directional consistency measurement (DCM) is defined to measure the stability of neighboring pixels’ alignment degree, which can distinguish the linear structure’s random noise pattern. Thirdly, by combiningDCMwith weighted grayscale correlation in the direction of estimation, the correlated signal is enhanced and the uncorrelated background noise is suppressed. Finally, parallel acceleration based on graphics processing unit (GPU) is implemented in computationally expensive part of the algorithm to achieve real-time application. The experimental results show that the proposed method achieves a detection probability exceeding 93% at SNR ≥ 2.0 with a false alarm probability below 9%.