Abstract:

The evolution process of dune morphology records the history of near-surface wind conditions and environmental evolution， but its characteristic research has been limited by the inefficient and high cost of extracting large-scale dune crest lines. For this reason， this paper builds a U-Net model based on the deep convolutional neural network for batch and high-precision extraction. In order to obtain the best extraction results， the enhancement technology in data preprocessing， random neurons inactivation， batch normalization and transfer learning technology have been applied to the training and parameters updating， making the prediction accuracy of the model higher. The results show that the model and various strategies used in this paper can efficiently and accurately identify the dune crest lines in remote sensing images. In addition， through the application study of the orientation of the dune crest lines extracted from the trained model， we can find the shift of the dune crest lines orientation has a good correspondence with the change of near-surface wind regime， and it is confirmed that the U-Net model can be effectively used in regional dune orientation analysis.

Key words: deep convolutional neural network, U-Net, dune crest lines, near-surface wind regime