Research on Automatic Identification Method of Terraces on the Loess Plateau Based on Deep Transfer Learning

Rapid, accurate extraction of terraces from high-resolution images is great significance for promoting the application remote-sensing information in soil and water conservation planning monitoring. To solve problem how deep learning requires a large number labeled samples to achieve good accuracy, this article proposes an automatic identification method that can obtain high precision through small sample datasets. Firstly, terrace source model adapted multiple data sources trained based on WorldView-1 dataset. The be migrated other types terracing as pre-trained model. Secondly, problem, transfer pixel-level image proposed. Finally, insufficient boundary splicing traces during prediction, strategy ignoring edges proposed, prediction constructed further improve accuracy identification. In paper, three regions outside area are randomly selected, OA, F1 score, MIoU averages reach 93.12%, 91.40%, 89.90%, respectively. experimental results show method, learning, accurately extract terraced field surfaces segment boundaries.