• 首页 | 主办单位 | 期刊简介 | 编委会 | 作者指南 | 刊物订阅 | 下载中心 | 联系我们 | English | 期刊界
引用本文:赵红伟,陈仲新,刘佳.深度学习方法在作物遥感分类中的应用和挑战[J].中国农业资源与区划,2020,41(2):35~49
【打印本页】   【下载PDF全文】   查看/发表评论  【EndNote】   【RefMan】   【BibTex】
←前一篇|后一篇→ 过刊浏览    高级检索
本文已被:浏览 101次   下载 0 本文二维码信息
码上扫一扫!
分享到: 微信 更多
深度学习方法在作物遥感分类中的应用和挑战
赵红伟1, 2陈仲新1, 2, 3※刘佳1, 2
1.中国农业科学院农业资源与农业区划研究所,北京100081; 2.农业农村部农业遥感重点实验室,北京100081; 3.联合国粮农组织信息技术部门(FAO),罗马00153
摘要:
[目的]准确估算作物的面积和分布对粮食安全至关重要。与传统的机器学习方法相比,深度学习具有多种优势,如端到端训练、可迁移性。为有效利用高时空数据进行作物识别提供了新的机遇。已有多种模型被应用于作物分类任务中,针对不同的分类任务,如何有效地选择模型,并对其进行训练和使用已成为关键问题。[方法]文章回顾了利用深度学习模型对作物分类的主要研究。N维卷积神经网络(N-D CNN)(N=1、2、3)和递归神经网络(RNN)已被有效用于作物分类任务。长短期记忆RNN(LSTM RNN)和门控循环单元RNN(GRU RNN)是RNN的变体,解决了随着时间序列增加RNN出现的梯度消失或爆炸问题。此外,还有研究使用CNN和RNN(我们称为RCNN)的混合模型对作物进行分类。该文首先阐述了使用深度学习方法进行作物制图的背景和意义,并介绍了CNN和RNN模型结构。然后回顾了一些典型的研究,包括模型的结构、遥感数据源、数据处理方法和分类精度。最后,总结了使用深度学习方法进行作物分类的挑战以及现有解决方案的局限性。[结果](1)1 D CNN可用于提取时间特征,或时间+光谱特征,分类效果良好; 2 D CNN已被广泛应用于单时相数据的空间特征提取,分类精度依赖于数据源; 3 D CNN 应用较少,但具有很大的潜力,尤其是时间+空间维度的特征提取; (2)相同条件下(架构、数据源、研究区域、类别),LSTM RNN 和GRU RNN 分类效果通常高于普通RNN,而前两者的效果差距不大,但GRU RNN训练时间较短; (3)CNN+RNN混合模型(RCNN)用RNN比3 D CNN更适合提取时间特征。这主要是由于RNN建立了对序列数据的长期依赖,而3 D CNN卷积核是局部计算的。[结论]通过分析,认为深度学习技术是作物遥感分类的有效工具。此外,与其他模型相比,RCNN, 3 D CNN和GUR RNN具有更大的潜力。
关键词:  深度学习农作物分类CNNLSTM RNNGRU RNN
DOI:
分类号:
基金项目:中央级公益性科研院所基本科研业务费专项“基于深度学习和Sentinel 2A/B时间序列影像的作物早期制图研究——以河北省衡水市为例”(1610132020031); 农业农村部现代农业人才支撑计划(农业空间信息技术创新团队)项目(914-2)
DEEP LEARNING FOR CROP CLASSIFICATION OF REMOTE SENSING DATA: APPLICATIONS AND CHALLENGES
Zhao Hongwei1,2, Chen Zhongxin1,2,3※,Liu Jia1,2
1. Institute of Agricultural Resources and Regional Planning, Chinese Academy of Agricultural Sciences, Beijing 100081, China; 2. Key Laboratory of Agricultural Remote Sensing, Ministry of Agriculture and Rural Affairs, Beijing 100081, China; 3. Information Technology Division, Food and Agriculture Organization of the United Nations (FAO), Rome 00153
Abstract:
Accurate estimation of the area and distribution of crops is vital for food security. Compared to classical machine learning methods, deep learning has many advantages, such as end to end training and mobility, resulting into a new possibility of using high spatio temporal data efficiently. Several deep learning models have been used on different remote sensing imagery. However, the choice of the best deep learning model, its training and effective, use remains as a question yet to be answered. Thus, this paper reviews a large number of researches on crop classification using deep leaning models. Two approaches of deep learning models which had been efficiently used for crop classification tasks: N dimension convolutional neural network (N D CNN) (N=1, 2, 3) and recurrent neural network (RNN). Long short term memory RNN (LSTM RNN) and gated recurrent unit RNN (GRU RNN) were variants of RNN that could solve problems of gradient disappearance or explosion seen due to increasing time series. In addition, some researches classified crops using the hybrid model of CNN and RNN (we referred it as RCNN). This paper firstly introduced the reason we used deep learning method to map crops and discuss each deep learning model. Then, well recognized researches were reviewed, including different structures of models, remote sensing data sources, data processing, and classification accuracies. Finally, challenges and limitations in crop classification using deep learning method were summarized. The results showed that(1)1 D CNN could be used to extract temporal features, or temporal and spectral features, attaining satisfactory results. 2 D CNN had been widely used for extracting spatial features of single phase data, and the classification accuracies deeply depend on the used data. 3 D CNN was less used, but had great potential, especially in extracting spatio temporal features.(2)When the structures of ordinary RNN, LSTM RNN and GRU RNN, the data sources, research regions, and crop categories were the same , LSTM RNN and GRU RNN showed better performance than ordinary RNN. Moreover, the accuracies of LSTM and GRU were similar, but the training time of GRU was shorter. (3) RCNN was more suitable for extracting temporal features than 3 D CNN, mainly because the RNN established long term dependence on the sequence data, and convolutional kernel was locally calculated. The analysis point out the deep learning as an effective tool for crop classification using remote sensing imagery. Although presenting great potential, researchers have propped solutions for challenges that need to be overcame. In addition, compared with other models, RCNN, 3 D CNN, and GUR have greater potential for crop classification.
Key words:  deep learning  crop classification  CNN  LSTM RNN  GRU RNN
版权所有:  您是本站第    位访问者
主管单位:中华人民共和国农业部 主办单位:中国农业科学院农业资源与农业区划研究所 中国农业资源与区划学会 地址:北京市中关村南大街12号
电话:010-82109647 82108697 电子邮件:quhuabjb@caas.cn
技术支持:北京勤云科技发展有限公司  京ICP备11039015号