KMS XINJIANG INSTITUTE OF ECOLOGY AND GEOGRAPHY，CAS
| 基于时序光谱重构的卷积神经网络遥感农作物分类与应用 | |
| Alternative Title | Remote Sensing Crop Classification and Application Based on Time Series Spectral Reconstruction using CNN |
| 冯齐心 | |
| Subtype | 硕士 |
| Thesis Advisor | 杨辽 |
| 2019-06-30 | |
| Degree Grantor | 中国科学院大学 |
| Place of Conferral | 北京 |
| Degree Discipline | 工程硕士 |
| Keyword | 农作物分类 卷积神经网络 遥感 时间序列 特征提取 Crop classification Convolutional neural network Remote sensing Time series Feature extraction |
| Abstract | 精准农业是世界各国农业现代化的共同选择。 随着遥感信息技术的快速发展， 遥感在农作物分类与种植结构调查方面也取得了长足进步， 但仍然存在以下问题： 一是目前的遥感农作物分类法均需要较多的人工干预， 自动化水平不高， 二是由于数据源与分类方法的限制，分类精度有待进一步提升。近年来， 遥感的交叉学科——计算机科学与技术在深度学习领域取得了丰硕成果， 尤其在图片的分类与识别领域，卷积神经网络可自动从图片中提取深层特征，完成高精度的图片分类与识别， 相比传统的手工设计特征提取器或机器学习算法有较大的精度优势， 且自动化水平更高， 可为解决遥感农作物分类中精度有待提高、 难以自动化等问题提供新思路。因此本文采用时间序列的Sentinel-2A 数据， 以新疆沙湾县为实验区， 开展了基于卷积神经网络的农作物分 类 方 法 研 究 ， 并 提 出 了 基 于 时 序 光 谱 重 构 的 卷 积 网 络 作 物 分 类 法(TSMI+CNN)。 该方法为充分利用时间序列数据中丰富的多光谱与光谱变化信息， 以地面像元为基本单元， 以时间维为纵轴、光谱维为横轴， 对时间序列多光谱进行了重构，得到了每个地面像元基元的时序光谱图， 再采用 Adam 梯度下降法与 Dropout 40%连接率优化后的卷积网络对时序光谱图进行分类。 为验证该方法的实际分类效果，本文采用了时间序列多光谱+随机森林、 时间序列NDVI+随机森林、 时间序列 NDVI+卷积神经网络等方法进行分类对比研究，结果表明：(1) 该方法总体分类精度达到了 95.12%， 高于时间序列多光谱+随机森林(88.58%)、时间序列 NDVI+随机森林(90.25%)、时间序列 NDVI+卷积神经网络(91.79%)等对照实验组， 表明该分类法可有效提高遥感农作物分类的精度。(2) 对于―异物同谱‖情况明显，在各对照组中混淆相对严重的春玉米与番茄，TSMI+CNN 分类法的 F1-score 综合精度分别达到了 95.9%和 89.9%，相比其他对比试验组中的最高精度的分别提高了 8.2%与 8.3%。 说明该方法可有效提高光谱相似性高作物的分类精度。(3) 本研究中各对照组均有不同程度的边界轮廓线模糊现象，不利于农作物的地块边界提取。 另外， TSM+RF 分类法地块内部不均匀，出现了较多椒盐噪声。TSMI+CNN 分类法的制图结果边界清晰，地块内部均质且基本无椒盐噪声，可完成高质量的农作物遥感制图， 能为基元像元法的遥感制图提供方法借鉴。(4) 基于特征提取的传统分类方法涉及复杂的特征提取与重建工作，TSMI+CNN 分类法通过对时间序列多光谱的时间维与光谱维重新排列， 采用卷积网络完成了研究区内各农作物的提取， 不涉及复杂特征提取与重建，可有效提高遥感农作物分类的自动化水平。 |
| Other Abstract | Precision agriculture is the common choice for agricultural modernization in allcountries of the world. With the rapid development of remote sensing informationtechnology, remote sensing has made great progress in crop classification, but thefollowing problems still exist: First, most of the current remote sensing cropclassification methods require lots of manual intervention, which is difficult toautomate. Second, Due to the limitations of data sources and classification methods,the classification accuracy needs to be further improved.In recent years, the interdisciplinary of remote sensing, computer science andtechnology has made a series of major breakthroughs in the field of deep learning.Especially in the field of image classification and recognition, convolutional neuralnetwork can automatically extract deep features from pictures and achievehigh-precision classification result. Compared with traditional machine learningalgorithms, convolutional neural network has greater precision advantages andhigher automation level, which can provide reference for solving the problems incrop classification. With these merits of CNN, this paper proposes a cropclassification method based on time-series spectral reconstruction and convolutionalneural network (TSMI+CNN). In order to make full use of the rich cropphenological information and multi-spectral information in time-series data, firstly,TSMI+CNN method translates every ground pixel's time-series multispectral into an×m image, and then uses CNN which optimized by Adam gradient descent anddropout(40%) neural connection method to classify time-series images. In order toverify the actual classification effect of TSMI+CNN method, this paper take theShawan of Xinjiang as the experimental area, uses TSM+RF, TSNDVI+RF, andTSNDVI+CNN method to compare with TSMI+CNN method. The results showthat:(1) The overall classification accuracy of TSMI+CNN method reches 95.12%,which is higher than TSM+RF(88.58%), TSNDVI+RF(90.25%), TSNDVI+CNN(91.79%) methods. Which shows that the TSMI+CNN method can effectivelyimprove the accuracy of crop classification.(2) For spring corn and tomato that with high spectral similarity, the F1-score ofTSMI+CNN method reaches 95.9% and 89.9% respectively, which is significantlyimproved compared with the control groups. Indicating that TSMI+CNN methodcan effectively distinguish crops with high spectral similarity.(3) TSMI+CNN mapping results show that the boundary of the cropclassification map is clear, the interior of the plot is uniform, and there is no obvious"pepper and salt" noise. Showing that TSMI+CNN classification method canachieve a high crop mapping quality, and has good practical application value.(4) The traditional classification method based on feature extraction involvescomplex feature extraction and reconstruction. However, without involving complexfeatures extraction and reconstruction, TSMI+CNN method successfully classifiedcrops in experimental area, meaning that TSMI+CNN method can effectivelyimprove the automation level of crop classification. |
| Subject Area | 测绘工程 |
| Language | 中文 |
| Document Type | 学位论文 |
| Identifier | http://ir.xjlas.org/handle/365004/15325 |
| Collection | 中国科学院新疆生态与地理研究所 研究系统 |
| Affiliation | 中国科学院新疆生态与地理研究所 |
| First Author Affilication | 中国科学院新疆生态与地理研究所 |
| Recommended Citation GB/T 7714 | 冯齐心. 基于时序光谱重构的卷积神经网络遥感农作物分类与应用[D]. 北京. 中国科学院大学,2019. |
| Files in This Item: | There are no files associated with this item. | |||||
Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.
Edit Comment