KMS XINJIANG INSTITUTE OF ECOLOGY AND GEOGRAPHY，CAS
| 基于深度学习的努库斯灌区耕地信息遥感提取 | |
| Alternative Title | Remote sensing extraction of cultivated land information in Nukus irrigation district based on deep learning |
| 杨志坚 | |
| Subtype | 硕士 |
| Thesis Advisor | 陈曦 |
| 2020-06-30 | |
| Degree Grantor | 中国科学院大学 |
| Place of Conferral | 北京 |
| Degree Discipline | 工程硕士 |
| Keyword | 深度学习 耕地提取 农作物分类 时间序列 多特征融合 Deep learning Cultivated land extraction Crop classification Time series Multi-feature fusion |
| Abstract | 耕地边界信息和农作物空间分布信息是进行农业资源监测和种植结构调整与优化的重要基础数据。 随着遥感信息智能提取技术的进步， 利用遥感影像开展耕地信息提取研究取得了快速发展， 但在地块破碎、 分布零散和种植结构相对复杂的区域提取效果欠佳， 另外提取方法也存在局限性。 近年来， 深度学习网络因为能挖掘出时间序列影像中更深层次特征信息， 在遥感图像识别和分类领域取得了重大突破， 弥补了传统方法的不足。鉴于此， 本文选取时间序列 Sentinel-2 数据， 以中亚努库斯灌区为实验区，结合实地采样数据， 开展了基于 U-Net 网络的耕地地块分割和基于多特征信息融合的 CNN 农作物分类研究， 实现了研究区内耕地信息的自动提取。 其中 U-Net网络通过卷积操作能提取到图像更多细节特征， 在反卷积中又能还原图像的原始特征， 并保留空间位置信息， 较好地满足了耕地边界提取的需求。 多特征信息融合方法则充分利用了作物的“光谱-时相-空间” 特征， 它以作物地面像元基元为最小分类单元， 将每个像元的三种特征值融合成一张图像， 每个图像包含了该像元的 80 个特征值， 卷积神经网络根据同种作物特征值图相似性较高， 不同作物特征值图相似性差异较大的特点进行分类， 并用不同样本特征组对分类结果精度对比验证。 结果表明：（ 1） U-Net 网络提取到实验区耕地面积 23530.46km2， 总体分割精度达到93.66%， 相比于 FCN 和 SegNet 对照组分别提高了 5.08%和 9.27%， 说明深度学习方法可以有效应用于中亚干旱区耕地边界提取， 精度高于传统分割方法。（2） 作物的“光谱-时相-空间” 特征信息进行融合， 在卷积神经网络中可取得较好的分类效果， 作物的总体分类精度达到 95.59%， 相比于时序反射率值+CNN 和单期多特征+CNN 对照组分别提高了 4.24%和 5.91%， 说明该分类方法可以有效提高干旱区农作物分类精度， 具有一定创新性。（3） 对整个努库斯灌区进行作物分类， 提取到棉花种植面积 9706.37km2，水稻面积 4651.21km2， 冬小麦面积 3710.97km2， 玉米面积 1413.06km2， 其分类一致性较好， 表明该方法适用于大规模、 大面积的农业遥感监测。 同时， 深度学习方法中的 U-Net 网络和 CNN 网络只需要少量的样本即可实现模型的训练， 可节省大量资源。（4） 基于耕地地块进行的耕地类型（作物类型） 分类首先剔除了地块以外的地类信息， 使所有作物分类过程均在纯净地块基础上进行， 可有效提高分类精度， 也可为其它地物监测提供思路， 比如： 林业方面的树种提取与分类。（5） 实验利用两种深度学习网络进行了快速、 准确地耕地信息提取， 相比于传统分类方法， 该过程不涉及复杂特征提取工作， 可以有效提高农业遥感耕地信息提取的自动化水平， 且训练好的模型可重复运用， 保障了农业监测的实时性。 |
| Other Abstract | Cultivated land boundary information and crop spatial distribution informationare important basic data for monitoring agricultural resources and adjusting andoptimizing planting structures. With the advancement of intelligent extractiontechnology of remote sensing information, research on the extraction of cultivatedland information has achieved rapid development by remote sensing imagery. Inrecent years, deep learning networks have demonstrated the ability to mine deep-levelfeatures in time series images, and have made significant breakthroughs in the field ofremote sensing image recognition and classification, making up for the shortcomingsof traditional methods.In view of this, this paper selects the time series Sentinel-2 data, takes the NukusIrrigation District of Central Asia as the experimental area, and combines fieldsampling data to carry out the segmentation of cultivated land based on U-Net andcrop classification based on CNN by multi-feature information fusion. The researchfinished the automatic extraction of cultivated land information in the study area.Among them, the U-Net can extract more detailed features of the image through theconvolution operation, and, restore the original features of the image in thedeconvolution, and retain the spatial position information, which satisfies the needs offarmland boundary extraction. The method of multi-feature information fusion makesfull use of the "spectral-temporal-spatial" feature of the crop. It uses the crop groundpixel primitives as the minimum classification unit, and fuses the three feature valuesof each pixel into an image. Each image contains 80 eigenvalues of this pixel. CNNclassifies the features based on the high similarity of eigenvalue map of the same cropand the dissimilarity of eigenvalue maps of the different crop, and compare and verifythe accuracy of classification results with different sample feature groups. The resultsshow that:(1) The U-Net extracted cultivated area 23530.46km2 in the experimental area,and the overall segmentation accuracy reached 93.66%. Compared with the FCN andSegNet control group, it improved by 5.08% and 9.27%, indicating that the deeplearning method can be effectively used in extraction of cultivated land boundary inarid area of central Asia, and the accuracy is higher than traditional methods.(2) Fusion of the "spectral-temporal-space" feature information of the crop canachieve better classification results by CNN. The overall classification accuracy of thecrop reaches 95.59%, compared to the reflectance value of time series + CNN andmulti-feature of single-phase + CNN control group, it improved by 4.24% and 5.91%,indicating that the classification method can effectively improve the accuracy of cropclassification in arid area of central Asia, and it has some innovation.(3) Crop classification was performed for the entire Nukus Irrigation District, andcotton planting area of 9706.37km2, rice area of 4651.21km2, winter wheat area of3710.97km2, and corn area of 1413.06km2. The classification consistency is excellent,indicating that the method is suitable for monitoring of large-scale and large-areaareas by agricultural remote sensing. At the same time, U-Net and CNN networks ofmethods of deep learning only need a small number of samples to implement modeltraining, saving a lot of resources.(4) Cultivated land type (crop type) classification based on cultivated land plots,Firstly remove the information of land type except for the plots, so that all cropclassification processes are performed on the basis of pure plots, which can effectivelyimprove the classification accuracy and can also provides ideas for other land featuresmonitoring , such as: extraction and classification of tree species in forestry.(5) The experiment uses two deep learning networks to extract cultivated landinformation quickly and accurately. Compared with traditional classification methods,this process does not involve the extraction of complex features, which can effectivelyimprove the level of automation of information extraction of cultivated land byagricultural remote sensing. guarantee the real-time nature of agricultural monitoring. |
| Subject Area | 测绘工程 |
| Language | 中文 |
| Document Type | 学位论文 |
| Identifier | http://ir.xjlas.org/handle/365004/15463 |
| 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 | 杨志坚. 基于深度学习的努库斯灌区耕地信息遥感提取[D]. 北京. 中国科学院大学,2020. |
| 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