Remote Sensing Image Change Detection With Transformers

Abstract

Translate Article

Modern change detection (CD) has achieved remarkable success by the powerful\ndiscriminative ability of deep convolutions. However, high-resolution remote\nsensing CD remains challenging due to the complexity of objects in the scene.\nObjects with the same semantic concept may show distinct spectral\ncharacteristics at different times and spatial locations. Most recent CD\npipelines using pure convolutions are still struggling to relate long-range\nconcepts in space-time. Non-local self-attention approaches show promising\nperformance via modeling dense relations among pixels, yet are computationally\ninefficient. Here, we propose a bitemporal image transformer (BIT) to\nefficiently and effectively model contexts within the spatial-temporal domain.\nOur intuition is that the high-level concepts of the change of interest can be\nrepresented by a few visual words, i.e., semantic tokens. To achieve this, we\nexpress the bitemporal image into a few tokens, and use a transformer encoder\nto model contexts in the compact token-based space-time. The learned\ncontext-rich tokens are then feedback to the pixel-space for refining the\noriginal features via a transformer decoder. We incorporate BIT in a deep\nfeature differencing-based CD framework. Extensive experiments on three CD\ndatasets demonstrate the effectiveness and efficiency of the proposed method.\nNotably, our BIT-based model significantly outperforms the purely convolutional\nbaseline using only 3 times lower computational costs and model parameters.\nBased on a naive backbone (ResNet18) without sophisticated structures (e.g.,\nFPN, UNet), our model surpasses several state-of-the-art CD methods, including\nbetter than four recent attention-based methods in terms of efficiency and\naccuracy. Our code is available at https://github.com/justchenhao/BIT\\_CD.\n