Adaptive Multisensor Fusion for Remote Sensing Change Detection Using USASE

Binary change detection (BCD) in remote sensing has advanced, yet challenges remain in reducing feature redundancy and effectively utilizing difference information between dual-time images, which affects precision in identifying change areas. In addition, the effective fusion of multisensor data types limits adaptability and accuracy in change detection (CD) models. This article presents the ultralightweight semantic-aware spatial exchange (USASE) network, a three-encoder-three-decoder architecture designed for improved adaptability in multisensor data fusion. USASE integrates a micro convolutional unit (MCU) for reduced feature redundancy through pointwise and depthwise separable convolutions, while a temporal-aware feature aggregation module (TAFAM) captures global semantic relationships to enhance detection precision across sensor types. An adaptive weighting mechanism further optimizes dual-time image accuracy in multisource data fusion. Tested on the SYSU-CD, LEVIR-CD, and DSIFN datasets, USASE achieves the ${F}1$ -scores of 83.12%, 90.72%, and 81.34%, respectively, outperforming several baselines in accuracy, efficiency, and computational cost. This study highlights USASEs potential as a robust, real-time solution for dynamic and complex remote sensing applications.