U-net深度学习神经网络结合面向对象的城市森林高分影像信息提取

U-net Neural Network Combined with Object-oriented Urban Forest High-score Image Information Extraction

城市森林结构复杂、分布破碎,采用高分遥感数据,通过深度学习等智能机器学习算法精准监测提取城市森林信息,是城市森林资源智能监测管理的关键性基础环节。本文以杭州市余杭区部分城区WorldView-3高分卫星遥感影像为数据源,采用改进的U-net深度学习神经网络,并结合面向对象多尺度分割方法,研究城市森林智能精准提取。首先,通过大量的训练数据获得最佳模型参数;其次,使用U-net网络进行语义分割得到分类结果图;最后,结合面向对象最优分割修正深度学习城市森林提取结果,从而最终得到城市森林提取结果。研究表明,(1)基于改进的U-net深度学习神经网络得到的城市森林总体分类精度达90.50%,Kappa系数为0.886;(2)经面向对象分割对U-net深度学习神经网络结果中的"椒盐现象"及边界地物错分现象进行修正后,分类总精度提高到93.83%,Kappa系数提高到0.9295。因此,U-net网络模型结合面向对象方法可以有效地改善遥感目标识别及地物分类的效果,保证城市碎片化植被提取与植被区域边界的准确性,从而提高城市森林植被提取精度。

Considering the complexity and fragmentation of urban forest structure,using high-resolution remote sensing data and intelligent machine learning algorithms such as deep learning to accurately extract urban forest information is crucial for intelligent monitoring and urban forest resources management.This paper used WorldView-3 highresolution satellite imagery of urban areas in Yuhang District,Hangzhou and an improved U-net deep learning neural network combined with an object-oriented multi-scale segmentation algorithm developed to extract urban forests classification information.Firstly,through a large amount of training data,it obtained optimized parameters.And then,the U-net network is used to get the classi cation map.Finally,it carried out an objectoriented multi-scale segmentation algorithm to modify the urban forest extraction results to obtain accurate urban forest information.The research showed that:(1)the overall classi cation accuracy of urban forest based on the improved U-net deep learning neural network is 90.50%,and the Kappa coef cient is 0.886;(2)analyzing with an object-oriented multi-scale segmentation algorithm,it h as corrected the salt and pepper noise and errors of boundar y features in the U-net deep learning neural network’s results have.The total classi cation accuracy increased to 93.83%,and the Kappa coef cient rose to 0.9295.Therefore,the U-net ne twork model combined with object-oriented multi-scale segmentation algorithms can effectively improve remote sensing vegetation classi cation precision.