Objectives  In recent years, unmanned aerial vehicle has been widely used and their navigation and positioning rely heavily onglobal navigation satellite system(GNSS). In the case of GNSS rejection, visual navigation and positioning technology can compensate for this problem, but the technique will also fail to adapt if the approximate location of unmanned aerial vehicle cannot be estimated.To cope with this problem, we propose a reference satellite image retrieval method that aggregates deep learning features to determine the range of unmanned aerial vehicle image adaption region, which can provide reference for the following unmanned aerial vehicle absolute positioning.

  Methods  Firstly, the pre-trained deep learning model is used to extract local convolution features of unmanned aerial vehicle images and satellite images.Secondly, the local aggregation descriptor vector is used to generate the global expression of the images. Finally, the global feature of the image is used to perform similarity retrieval and post-processing method of matching precisely and reranking the retrieval results is used, which further improves the retrieval accuracy.A new satellite reference image data set for absolute positioning of unmanned aerial vehicle is designed and tested.

  Results  When the queried unmanned aerial vehicle image is similar to the satellite image season in the database, the accuracy of the top 50 candidate images can reach 87.50% using the proposed features for retrieval. Combined with the refined matching re-ranking, the accuracy of the first candidate image can reach up to 76.07%, which satisfies general navigation and positioning applications.

  Conclusions  Although the global descriptor based on deep feature aggregation can effectively represent the images of texture-obvious regions, it is not strong in representing the images of texture-lacking regions and its overlap range between images is high when retrieval is performed.Therefore, the efficiency problem of retrieval and the image feature representation of texture-sparse regions are the directions that need further research.