{"title":"基于密集连接卷积神经网络和集成分类器的遥感图像场景分类","authors":"Q. Cheng, Yuan Xu, Peng Fu, Jinling Li, Wen Wang, Y. Ren","doi":"10.14358/PERS.87.3.295","DOIUrl":null,"url":null,"abstract":"Deep learning techniques, especially convolutional neural networks, have boosted performance in analyzing and understanding remotely sensed images to a great extent. However, existing scene-classification methods generally neglect local and spatial information that is vital to scene\n classification of remotely sensed images. In this study, a method of scene classification for remotely sensed images based on pretrained densely connected convolutional neural networks combined with an ensemble classifier is proposed to tackle the under-utilization of local and spatial information\n for image classification. Specifically, we first exploit the pretrained DenseNet and fine-tuned it to release its potential in remote-sensing image feature representation. Second, a spatial-pyramid structure and an improved Fisher-vector coding strategy are leveraged to further strengthen\n representation capability and the robustness of the feature map captured from convolutional layers. Then we integrate an ensemble classifier in our network architecture considering that lower attention to feature descriptors. Extensive experiments are conducted, and the proposed method achieves\n superior performance on UC Merced, AID, and NWPU-RESISC45 data sets.","PeriodicalId":49702,"journal":{"name":"Photogrammetric Engineering and Remote Sensing","volume":null,"pages":null},"PeriodicalIF":1.0000,"publicationDate":"2021-04-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"2","resultStr":"{\"title\":\"Scene Classification of Remotely Sensed Images via Densely Connected Convolutional Neural Networks and an Ensemble Classifier\",\"authors\":\"Q. Cheng, Yuan Xu, Peng Fu, Jinling Li, Wen Wang, Y. Ren\",\"doi\":\"10.14358/PERS.87.3.295\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"Deep learning techniques, especially convolutional neural networks, have boosted performance in analyzing and understanding remotely sensed images to a great extent. However, existing scene-classification methods generally neglect local and spatial information that is vital to scene\\n classification of remotely sensed images. In this study, a method of scene classification for remotely sensed images based on pretrained densely connected convolutional neural networks combined with an ensemble classifier is proposed to tackle the under-utilization of local and spatial information\\n for image classification. Specifically, we first exploit the pretrained DenseNet and fine-tuned it to release its potential in remote-sensing image feature representation. Second, a spatial-pyramid structure and an improved Fisher-vector coding strategy are leveraged to further strengthen\\n representation capability and the robustness of the feature map captured from convolutional layers. Then we integrate an ensemble classifier in our network architecture considering that lower attention to feature descriptors. Extensive experiments are conducted, and the proposed method achieves\\n superior performance on UC Merced, AID, and NWPU-RESISC45 data sets.\",\"PeriodicalId\":49702,\"journal\":{\"name\":\"Photogrammetric Engineering and Remote Sensing\",\"volume\":null,\"pages\":null},\"PeriodicalIF\":1.0000,\"publicationDate\":\"2021-04-01\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"2\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Photogrammetric Engineering and Remote Sensing\",\"FirstCategoryId\":\"89\",\"ListUrlMain\":\"https://doi.org/10.14358/PERS.87.3.295\",\"RegionNum\":4,\"RegionCategory\":\"地球科学\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q4\",\"JCRName\":\"GEOGRAPHY, PHYSICAL\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Photogrammetric Engineering and Remote Sensing","FirstCategoryId":"89","ListUrlMain":"https://doi.org/10.14358/PERS.87.3.295","RegionNum":4,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q4","JCRName":"GEOGRAPHY, PHYSICAL","Score":null,"Total":0}
Scene Classification of Remotely Sensed Images via Densely Connected Convolutional Neural Networks and an Ensemble Classifier
Deep learning techniques, especially convolutional neural networks, have boosted performance in analyzing and understanding remotely sensed images to a great extent. However, existing scene-classification methods generally neglect local and spatial information that is vital to scene
classification of remotely sensed images. In this study, a method of scene classification for remotely sensed images based on pretrained densely connected convolutional neural networks combined with an ensemble classifier is proposed to tackle the under-utilization of local and spatial information
for image classification. Specifically, we first exploit the pretrained DenseNet and fine-tuned it to release its potential in remote-sensing image feature representation. Second, a spatial-pyramid structure and an improved Fisher-vector coding strategy are leveraged to further strengthen
representation capability and the robustness of the feature map captured from convolutional layers. Then we integrate an ensemble classifier in our network architecture considering that lower attention to feature descriptors. Extensive experiments are conducted, and the proposed method achieves
superior performance on UC Merced, AID, and NWPU-RESISC45 data sets.
期刊介绍:
Photogrammetric Engineering & Remote Sensing commonly referred to as PE&RS, is the official journal of imaging and geospatial information science and technology. Included in the journal on a regular basis are highlight articles such as the popular columns “Grids & Datums” and “Mapping Matters” and peer reviewed technical papers.
We publish thousands of documents, reports, codes, and informational articles in and about the industries relating to Geospatial Sciences, Remote Sensing, Photogrammetry and other imaging sciences.