To deal with the difficulty of remote sensing image target detection in foggy weather,an improved method based on Mask R-CNN is proposed.The defogging algorithm is added on the basis of Mask R-CNN, which improves the mean average precision by 18.71% in foggy weather and effectively improves the target detection effect in fog.In order to further improve the mean average precision of multi-scale target detection in remote sensing images,a recurrent neural network based on optimal feature combination is used to replace the structure of feature pyramid,which reduces the loss of feature information in the transmission process.The Region Proposal Network (RPN) is redesigned to generate the sizes of the candidate box,and Soft-NMS is used to screen the candidate box to reduce the regression error of the candidate box.Experimental analysis shows that the mean average precision and recall of the improved algorithm are improved by 5.37% and 6.37% respectively.