城市河网的低空遥感影像全卷积神经网络水质等级分类

Water Quality Classification of Low-altitude Remote Sensing Image of Urban River Network Based on Fully Convolutional Neural Network

摘要:

提出一种基于深度学习的图像像素级标注算法。通过数据预处理、数据集建立、全卷积神经网络设计和训练流程，实现水体的水质等级分类及像素级标注。使用上海市嘉定区某区域和上海市宝山区杨行镇某区域的无人机低空遥感影像对该算法进行了验证，平均水质等级分类精度分别达到了87.96%和77.57%。

Abstract:

This paper presents a dense semantic labeling algorithm based on deep learning. Firstly, low-altitude remote sensing data are acquired and preprocessed,and a data set for deep learning is built. Secondly, a fully convolutional neural network is designed and trained on the data set. Finally, the trained neural network is used to predict water quality level for each pixel in the remote sensing images. The algorithm is verified on image data acquired by unmanned areial vehicle（UAV） through low-altitude remote sensing in Jiading District and Baoshan District, Shanghai. Average classification accuracy achieves 87.96% and 77.57%, respectively.