Abstract: Aiming at the problem that the accuracy of remote sensing image retrieval is reduced due to the incomplete description of remote sensing images by a single type of features, this paper proposes a network model VDFF-Net that can fuse the underlying visual features with the deeper features. The model sliced the remote sensing images to get the small targets in them, and inputted them into the VGG-16 network model to get the deep features, as well as enhanced the significant information in the deep features by using the attention mechanism module. At the meantime, visual features were extracted from the slice image. Principal component analysis was used to reduce the dimension of the two types of features, the features were weighted and fused, and the cosine similarity was used to calculate the cosine distance between the query image and the image to be retrieved, which returned close distance as the retrieval result. The experiment was conducted in AID and NWPU-RESISC45 datasets. The results show that the fusion features extracted by VDFF-Net network model have better description of remote sensing image contents, which can effectively improve the accuracy of remote sensing image retrieval.