关键词: 梯田识别, 无人机影像分类, 深度学习, 语义分割

Abstract:

Terraces are the main water and soil conservation projects on sloping farmland. Accurate extraction of terrace information is important for monitoring and evaluating soil and water conservation. Aiming at the problem of terrace feature automatic learning in the research on UAV（Unmanned Aerial Vehicle） remote sensing terrace identification, this paper makes a set of terrace orthophoto sample sets with pixel level annotation and designs a terrace identification method combining the FCN-8s（Fully Convolutional Networks, FCN） model with the DenseCRF（Conditional Random Field, CRF） model. The experimental results show that the mean overall accuracy, F1 score and Kappa coefficient of this method in identification of ridged terraces, intensive horizontal terraces and irregular terraces is 86.85%, 87.28%, and 80.41%. Compared with other methods, the identification effect of this paper is better. This method is applicable to the field of UAV remote sensing terrace identification, and it is an accurate and effective identification method.

Key words: terrace identification, UAV image classification, deep learning, semantic segmentation