【摘要】 针对不同时相的多源遥感影像存在的空间异质性问题，本文对全变网络模型FTN (Fully Transformer Network)进行改进，提出一种端到端、基于滑窗式特征增强和卷积注意力混合机制的倒塌建筑物变化检测网络模型SWSACNet (Sliding-Window-Shift Attention Convolution mix Network)。SWSACNet基于FTN的模型框架，使用ACmix (Attention Convolution mix)高效识别多源影像对中的倒塌建筑物特征，并通过滑窗相似度特征匹配减弱多源影像中位置偏差的影响。以2023年2月6日土耳其Mw7.8级地震为例，通过获取震前高分二号、Google影像和震后北京三号影像构建倒塌建筑物变化检测数据集，对SWSACNet、FTN等5种变化检测模型进行训练和震区倒塌建筑物提取测试。实验结果表明，SWSACNet识别精度F1 score达80.8%,mIoU为67.8%，均优于其他4类模型。SWSACNet在应用于Fevaipasa、Nurdagi和Islahiye 3个测试场景中，模型平均识别精度F1 score为60.84%，表明模型在泛化性能上有待提升。更多还原

【Abstract】 Objective Change detection networks based on deep learning are widely used in water monitoring and urban transformation.However, collapsed buildings, as a change objective, are rarely targeted for change detection networks. This study proposes an end-to-end collapsed building extraction model based on a change detection network including sliding-window feature enhancement and convolution attention mix mechanism, which is called Sliding-Window-Shift Attention Convolution mix Network(SWSACNet).Method SWSACNet is an improvement of the Fully Transformer Network(FTN). FTN is a network completely composed of Swin Transformer. It has a unique frame, which involves four parts: Siamese Feature Extraction(SFE), Deep Feature Enhancement(DFE), Progressive Change Prediction(PCP), and Deep Supervision(DS). By encoding and decoding the feature of change objects, FTN can discern the modifications in collapsed buildings across two temporal images and suppress irrelevant information. ACmix, a blend of convolution and attention mechanism, has demonstrated better performance than Swin Transformer in mainstream datasets. As a result of different sensors and platforms, the spatial heterogeneity of target features in different source remote sensing images will affect the accuracy of change detection. Concerning this problem, we designed a similarity sliding window to match the feature maps of two temporal images. Thus, we replace Swin Transformer with ACmix to extract and restore earthquake-damaged features efficiently in the phase of SFE and PCP. We then use the similarity sliding window to reduce misidentifications of collapsed buildings in different source image pairs before the phase of DFE. Result Taking the earthquake with 7.8 magnitude that occurred on February 6, 2023, in Turkey as an example, we established a building seismic damage change detection dataset, which consisted of pre-earthquake Gaofen-2, Google images, and post-earthquake Beijing-3 images, and collapsed buildings were extracted based on the SWSACNet, FTN, STANet based on the Siamese self-attention mechanism, DASNet based on a dualattention fully-convolutional neural network, and the conventional fully-convolutional early fusion FC-EF network. The experimental results showed that SWSACNet achieved the highest accuracy with F1 score of 80.8% and mIoU of 67.8%. The ablation experiments of the improved model indicated that SWSACNet obtained the highest precision among three structural combinations. By smoothing the BJ-3image with high spatial resolution to align the gradient change rates of image pairs, we generated a new dataset and used it to retrain the five models. We found that the precision of the five retrained models increased by at least 1%, illustrating that appropriately narrowing the gap of gradient change rate of image pairs is an effective preprocessing technique for models to recognize collapsed buildings. Finally, by applying SWSACNet to three different data combinations covering Fevaipasa, Nurdagi, and Islahiye area, the results showed that it achieved an average F1 score of 60.84%. Conclusion The application of SWSACNet indicated that the model needs rich pre-and post-earthquake training dataset and structural improvement to enhance its generalization.更多还原