当前位置: X-MOL 学术J. Indian Soc. Remote Sens. › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
Evaluation of Convolutional Neural Networks for Urban Mapping Using Satellite Images
Journal of the Indian Society of Remote Sensing ( IF 2.5 ) Pub Date : 2021-06-01 , DOI: 10.1007/s12524-021-01382-x
Mina Mohammadi , Alireza Sharifi

Very high resolution (VHR) satellite imagery and image processing algorithms allow for the development of remote sensing applications including multi-temporal classification, tracking of specific targets, multimedia data integration, ecosystem processes analysis, and land cover/land use (LULC) mapping. Classification algorithms are the primary source to generate LCLU maps. Since texture information is essential to generate LULC maps from VHR images, the object-based classification methods should be used instead of pixel-based methods. Also, in urban mapping, it is vital to select the appropriate classifier according to the type of land covers. Recently, in addition to machine learning algorithms, deep learning methods have also been used to classify VHR images. In this study, we compare the accuracy of convolutional neural network (CNN) algorithm with some machine learning methods, for classification of Pleiades satellite image with 50 cm spatial resolution. The results showed CNN algorithm has the highest classification accuracy when the training samples are increased. However, the difference between the classification accuracy of the CNN and relevance vector machine (RVM) models is not that significant so that one could use a more straightforward method with less training data rather than a complicated one with large volumes of data.



中文翻译:

使用卫星图像评估用于城市测绘的卷积神经网络

超高分辨率 (VHR) 卫星图像和图像处理算法允许开发遥感应用,包括多时间分类、特定目标跟踪、多媒体数据集成、生态系统过程分析和土地覆盖/土地利用 (LULC) 制图。分类算法是生成 LCLU 地图的主要来源。由于纹理信息对于从 VHR 图像生成 LULC 地图至关重要,因此应使用基于对象的分类方法而不是基于像素的方法。此外,在城市测绘中,根据土地覆盖类型选择合适的分类器至关重要。最近,除了机器学习算法,深度学习方法也被用于对 VHR 图像进行分类。在这项研究中,我们将卷积神经网络 (CNN) 算法与一些机器学习方法的准确性进行了比较,以对具有 50 厘米空间分辨率的昴宿星卫星图像进行分类。结果表明,当训练样本增加时,CNN算法的分类准确率最高。然而,CNN 和相关向量机 (RVM) 模型的分类精度之间的差异并不那么显着,因此人们可以使用训练数据较少的更直接的方法,而不是具有大量数据的复杂方法。

更新日期:2021-06-02
down
wechat
bug