（参加第18届国际矿山测量大会研究生论坛）

    Change detection is an important task in the field of surveying and remote sensing, and plays a key role in many practical applications. Such as urban planning, environmental monitoring, natural resource management, disaster assessment, etc. Remote sensing imagery is characterized by large-area coverage, multispectral information, high resolution, and time-series information, making it a common data source for change detection. Traditional remote sensing image change detection methods are usually implemented by manually designing feature engineering. However, these methods have many limitations, such as the subjectivity and uncertainty of feature selection and the complexity of model parameter setting, etc. In order to better realize the change detection of ground objects, based on the robustness and end-to-end characteristics of deep learning, many scholars at home and abroad have carried out research on remote sensing image change detection with deep learning.

    Based on the double-branch structure and shared weight characteristics of Siamese neural network, it can be applied to remote sensing image change detection. This paper proposes a remote sensing image change detection method based on Siamese neural network. This method effectively utilizes the context information of remote sensing images, and the attention mechanism is integrated to extract the differences in remote sensing images more reasonably and accurately. Different from the existing deep learning methods, the method in this paper first uses VGG16 to extract the features of two images, maps the two input images to a weight sharing space, then fuses the extraction results of each scale, and then calculates the fusion The Euclidean distance between the results, on this basis, using the BAM (Bottleneck Attention Module) attention mechanism and the serial structure of the CEM (Bottleneck Attention Module) module for analysis, to obtain the difference of different sizes between the two input images part. Specifically, in the method of this paper, considering that the ground object changes in remote sensing images are usually regional, this paper introduces the CEM module to perceive the context information of each region of the image, so as to improve the model's ability to perceive the information of the different parts of the image. Considering the diversity of the area and color of the change area, a BAM attention mechanism is proposed to strengthen the model's attention to the change area, so as to solve the problem of missing extraction in the extraction results.

    In order to verify the effectiveness and reliability of our method, we conducted a series of experiments on public datasets CDD and LEVIR-CD, and compared our method with several other Siamese neural networks. The experimental results show that on the CDD dataset, the F1 value of the test set of this method is 0.9231, which is significantly better than other methods based on Siamese neural networks; on the LEVIR-CD dataset, the F1 value of the test set of this method is 0.8159, which is also more advantage. The above experimental results show that the method in this paper has better accuracy and robustness in remote sensing image change detection.