Abstract: Focusing on the large-scale jujube fields in the southern Xinjiang,this paper proposes a jujube orchard detection method based on a generalized deep transfer learning principle.From GF-6 satellite imagery,a jujube field dataset is made,and then it is augmented effectively.Grounded on a Faster R-CNN system,a multi-modally cooperative mode is used to realize the effective correlation and optimization reconstruction of the expanded dataset,and a transfer deep learning of detection and recognition model is thus carried out to improve the generalization ability of the detection and recognition of target object on jujube fields.The results show that the precision,recall and F1-score of the model algorithm reached 0.979,0.952 and 0.965,respectively.In the application tests,the average values of the three indexes are all more than 0.929,which could better than traditional detection method,and the overall classification accuracy and Kappa coefficient of this model method are 0.97 and 0.93,which are higher than the object-oriented nearest neighbor method,and effectively meet the requirements of high-efficient and accurate large-scale jujube orchard detection in the study area.Then it provides the basis for fine jujube orchard field management.

Key words: jujube orchard detection;Faster R-CNN;generalized transfer learning;data augmentation;GF-6