Remote sensing change detection aims to compare multi-temporal remote sensing images at the same location and identify significant as well as potential changes between them. Most of the related works focus on the chronological changes but perform poorly on anti-chronological detection. To avoid temporal effect, a common approach is to involve both chronological and anti-chronological data into datasets, but the model training time would be doubled simultaneously. Therefore, this paper proposes a 2-channel siamese network to ensure high accuracy as well as efficient training at the same time. Firstly, a symmetric model is constructed based on existing models to achieve fast training only using the original chronological datasets and to learn both chronological and anti-chronological features. Next, 2-channel siamese input model is designed to wrap the inputs for more robust feature extraction. Finally, attention mechanism is applied to further fuse and refine the extracted features. The proposed method is evaluated on the Onera Satellite Change Detection Sentinel-2 dataset. The Proposed model outperforms several existing models in terms of both accuracy and training validity. A further ablation study verifies the efficacy of proposed models.