| description abstract | As a large agricultural country, China’s timely and accurate extraction of rural buildings from high spatial resolution remote sensing images plays a crucial role in rural revitalization. With the public availability of the data set of buildings covering almost all the main structures in rural China published by China Scientific Data, this study proposes an improved building extraction method based on the above data set to address the problems of the irregularity of the rural building boundaries extraction and the neglect of the different recognition accuracies of the buildings with different main structures in the extraction process. Firstly, the VGG-Net network residual structure is introduced into the backbone feature extraction network of the U-Net model to extract deeper features; then, the attention mechanism is introduced to improve the extraction accuracy of rural buildings of different main structure types by adjusting their positions and numbers. Finally, the extraction results of the model with the addition of the attention mechanism are combined with the original image to form a four-channel image, which is again subjected to fine detection in order to improve the extraction accuracy of the building edges. The experimental results on the Chinese rural buildings data set show that the improved U-Net model outperforms other models in intersection over union (IOU), F1-score, precision, and recall accuracy evaluation metrics; further, the accuracy is improved in the extraction of buildings with different main structures, and the boundaries are more regularized, which can be applied in the extraction of buildings with different main structures. Meanwhile, the accuracies on GF-2, WHU data set, and Massachusetts data set are also improved, proving that the method proposed in this paper is robust and universal. | |
