Abstract: Aiming at the problem of  insufficient data volumes or low accuracy of labeled images in the field of remote sensing image change detection, which led to the model being unable  to fully learn features, and affected  the accuracy of detection, we proposed an improved  FCA-EF model based on the U-Net network. Firstly, the model was based on multi-head self-attention mechanisms and Transformer module of feedforward neural networks to establish encoding layers. Through long-distance skip connection mechanism, the  global features of the data were extracted in the encoding layer, achieving  information transfer between different layers. Secondly,  the model used convolutional neural network (CNN) module as the backbone to establish  decoding layers, extracted deep local features by using  the local perceptual characteristics of CNN module,  and fused the global features extracted by the encoder via long-distance skip connection mechanism to enhance the model’s ability to capture details and accuracy of  change detection. Thirdly, a new label filling and optimization method was proposed to address the problem of incomplete information representation in label image,  and its effectiveness was confirmed through ablation experiments. Finally, combined with the FCA-EF model and label filling method, the proposed method achieved excellent results inthe change detection of remote sensing images from Jilin-1 satellite. Compared with other classical models, the  overall accuracy, F₁ score, recall rate,  intersection over union (IoU) and other indicators were improved, effectively improving the accuracy of remote sensing image change detection.

Key words: remote sensing, change detection, FCA-EF model, label filling method