文章基本信息

标题：Semantic Segmentation for Remote Sensing based on RGB Images and Lidar Data using Model-Agnostic Meta-Learning and Partical Swarm Optimization
本地全文：下载
作者：Kai Zhang ; Yu Han ; Jian Chen 等
期刊名称：IFAC PapersOnLine
印刷版ISSN：2405-8963
出版年度：2020
卷号：53
期号：5
页码：397-402
DOI：10.1016/j.ifacol.2021.04.117
语种：English
出版社：Elsevier
摘要：AbstractUrban remote sensing has the problems that land cover categories are usually highly unbalanced and annotated samples are scarce, which brings great limitations to monitoring the change of urban coverage and periodically evaluating urban ecological conditions. Semantic segmentation is one of the main applications in urban remote sensing image analysis. Because the ground objects in remote sensing images have the characteristics of disordered distribution and irregular morphology. The classical semantic segmentation model based on deep learning U-Net cannot achieve high semantic segmentation accuracy for urban ground objects. This paper proposes to optimize the neural network structure and introduce lidar data to solve the above problems. In this paper, the Model-Agnostic Meta-Learning and fully convolutional neural networks are fused which be trained and tested by remote sensing images. It makes the training process into inner loop and outer loop. And Partical Swarm Optimization (PSO) is used to optimize the parameter updating process of neural network to further improve the test accuracy. This paper fuses Lidar data and remote sensing images, and comprehensively use the position and elevation information of Lidar data and the spectrum and texture information of remote sensing images to classify the ground features. The datasets used in this paper are RGB remote sensing images and Digital Elevation Model (DEM) images. The test accuracy of U-Net network optimized by MAML can be improved by 6%-7% under the same network parameters and training data sets. With the introduction of Lidar data, the accuracy of the test is increased by 3-5%. The experimental results show that the precision before and after PSO optimization is improved by 6%-9%, which verifies the idea in the paper.
关键词：KeywordsUrban remote sensingsemantic segmentationModel-Agnostic Meta-Learning (MAML)Partical Swarm Optimization (PSO)Lidar dataRGB remote sensing imagesDigital Elevation Model (DEM)