基于全卷积神经网络的灌区无人机正射影像渠系提取

Extraction of Irrigation Networks in Irrigation Area of UAV Orthophotos Based on Fully Convolutional Networks

摘要:

为快速准确获取灌区渠系分布信息，科学调配区域农业水资源、提高水资源利用率，通过基于全卷积神经网络（Fully convolutional networks，FCN）的语义分割模型进行渠系轮廓提取。利用无人机采集正射影像并进行标注，以VGG-19网络为基础，通过多尺度特征融合的方式实现FCN-8s结构，使用Tensorflow深度学习框架构建FCN渠系提取模型；对数据集进行数据增强，分割后放入FCN模型中训练、测试。实验结果显示，针对不同复杂程度的测试区域，FCN模型的提取准确度、完整度、精度均高于支持向量机方法和改进霍夫变换方法，均值分别为95.78%、92.29%、89.45%。结果表明，该方法能够实现灌区渠系轮廓的高精度提取，具有较好的泛化性和鲁棒性。

关键词:渠系;提取;全卷积神经网络;无人机;正射影像;语义分割

Abstract:

The distribution information of irrigation networks in irrigation area acquired quickly and accurately had great important research significance, especially in the scientific allocation of regional agricultural water resources and improvement of water resources utilization rate. The semantic segmentation model based on fully convolutional networks (FCN) was used to extract the irrigation networks contours. Firstly, the orthophotos collected by UAV were manually labeled. Based on the VGG-19 network, the FCN-8s structure was realized by multi-scale feature fusion, and the Tensorflow deep learning framework was used to construct the FCN irrigation networks extraction model. Secondly，the data sets were enhanced and segmented. Lastly, the data sets were put into the FCN model for training and testing. The experimental results showed that for the test areas with different complexities, the extraction precision, completion and accuracy of the FCN model were 95.78%, 92.29% and 89.45%, respectively, which were higher than the support vector machine (SVM) method and the revised Hough transform (RHT) method. The results showed that the method can achieve high-accuracy extraction of the irrigation networks contours in irrigation area, and had good generalization and robustness, which provided good technical support for further accurate irrigation in agriculture.

Key words:irrigation networks;extraction;fully convolutional networks;UAV;orthophoto;semantic segmentation