Current research in coseismic landslide identification typically follows a conventional supervised learning workflow, which involves dividing a seismic zone into 2 areas: one for training and the other for testing, or cropping the entire image and randomly splitting it into training and testing datasets. However, earthquakes often occur under varying terrain, geological, and vegetation cover conditions, leading to coseismic landslides that exhibit diverse shapes, spectral characteristics, and sizes. As a result, both CNN-based and Transformer-based models struggle to adapt to these variations. Additionally, models trained on a specific scale may face challenges in generalizing to other scales, making it difficult to accurately identify coseismic landslides of different magnitudes. Therefore, recognition of coseismic landslides across various seismic events is facing the predicament of low accuracy. To break through this predicament, this work suggests a novel Gated Dual-Stream Convolutional Neural Network (GDSNet) for cross-seismic-region recognition. This deep network comprises 2 branches and a feature fusion module. One branch, referred to as CPSConv, is designed to extract detailed landslide characteristics, including shapes, sizes, spectra, and textures, ensuring precise identification of small landslides. The other branch employs a gated convolution strategy to dynamically adjust feature weights and importance, thereby enhancing landslide features while suppressing background interference. The feature aggregation and fusion module integrates the features from both branches, significantly improving the recognition accuracy of coseismic landslides. The GDSNet has been successfully applied in 4 earthquake events: 2017 Ms 7.0 Jiuzhaigou earthquake, 2021 Mw 6.6 Iburi earthquake, 2021 Mw 7.2 Haiti earthquake, and 2022 Ms 6.1 Lushan earthquake.