WO2022214039A1 - 一种农业种植结构的遥感识别方法 - Google Patents

一种农业种植结构的遥感识别方法 Download PDF

Info

Publication number
WO2022214039A1
WO2022214039A1 PCT/CN2022/085640 CN2022085640W WO2022214039A1 WO 2022214039 A1 WO2022214039 A1 WO 2022214039A1 CN 2022085640 W CN2022085640 W CN 2022085640W WO 2022214039 A1 WO2022214039 A1 WO 2022214039A1
Authority
WO
WIPO (PCT)
Prior art keywords
sample set
agricultural
algorithm
training
classifier
Prior art date
Application number
PCT/CN2022/085640
Other languages
English (en)
French (fr)
Inventor
黄冠华
冯子怡
Original Assignee
中国农业大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 中国农业大学 filed Critical 中国农业大学
Publication of WO2022214039A1 publication Critical patent/WO2022214039A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/188—Vegetation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/088—Non-supervised learning, e.g. competitive learning

Definitions

  • the present disclosure belongs to the field of agricultural remote sensing, and relates to a remote sensing identification method for agricultural planting structures, which uses machine learning algorithms to extract crop planting structures, satisfies a variety of training data types, and obtains high-precision agricultural planting structure rasterized data after automatic training.
  • remote sensing can accurately obtain agricultural production information, and has become the main technical method to guide the transformation of traditional agriculture to information-based agriculture.
  • the use of remote sensing technology can realize crop identification and corresponding planting area estimation, crop growth monitoring and yield estimation, crop foliar index, chlorophyll content monitoring and nutrient diagnosis, agricultural land extraction and dynamic monitoring, etc. Therefore, the combination of remote sensing and geographic information technology has become a basic means to quickly collect and quantitatively analyze agricultural information and achieve scientific and rapid decision-making.
  • the identification of crop planting structure and the estimation of planting area are the basis for the adjustment of agricultural structure, and are also an important data source for the study of hydrological models and agricultural decision-making.
  • ELM Extreme Learning Machine
  • ELM Extreme Learning Machine
  • ELM The characteristic of ELM is that its learning process does not need to adjust the parameters of the hidden layer nodes, and the feature map from the input layer to the hidden layer can be random or artificially given. Since only the output weights need to be solved, ELM is essentially a linear-in-the-parameter model, and its learning process tends to converge at a global minimum.
  • the advantages of ELM in learning efficiency and generalization performance have been confirmed in many fields. In comparison with Back-Propagation (BP) and Support Vector Machine (SVM), the learning time of ELM significantly shortened with comparable learning accuracy.
  • BP Back-Propagation
  • SVM Support Vector Machine
  • ELM was originally designed for traditional classification and regression problems, but in subsequent research, its application range has been extended to almost all machine learning fields, including clustering and feature learning (representational learning), etc., and many variations have emerged.
  • ELM maintains the applicability of the universal approximation theorem of feedforward neural networks, and its learning strategies have brought many inspirations to the field of machine learning.
  • ELM cannot effectively extract deep abstract information of data. Therefore, the application of deep extreme learning machine is very important to solve the identification and application of planting structure under the condition of large data sparse samples.
  • the advantage of the supervised classification algorithm is that it can classify the pixels according to the training data samples, and determine the attribution of the pixels according to the characteristics of the samples.
  • the main defect of supervised classification is that the nature of the samples must be defined before classification, and the classification efficiency is low.
  • Non-supervised classification is to automatically discriminate and classify according to the similarity of pixels without samples, without manual intervention, and has a faster speed.
  • the classification effect is not as good as that of the supervised classification, and the unsupervised classification cannot distinguish the classification of the pixels, which is a clustering algorithm.
  • the existing classification algorithms are difficult to implement in large-scale high-precision images, and the classification efficiency of a single classification algorithm and classifier is low.
  • the purpose of the present disclosure is to provide a remote sensing identification method for agricultural planting structures, which firstly clusters similar pixels into a single category through the unsupervised algorithm k-means, and then uses the self-labeling algorithm to expand the sample set. Finally, the D-ELM algorithm is used to train and calculate these merged pixel categories. After training, all regions are classified and discriminated, which greatly improves the efficiency of the classifier and avoids the problem that the use of supervised classification algorithm is too complex in time and space, which makes the classification of large-area planting structures infeasible. At the same time, D-ELM hybrid classification Under the premise of ensuring the classification accuracy, the device can meet the mapping accuracy of agricultural remote sensing planting structure identification.
  • a remote sensing identification method for agricultural planting structures comprising the following steps:
  • Step S1 using the k-means unsupervised learning algorithm to segment the original satellite image:
  • step S1.2 Calculate the distance between the unclustered pixel points in the original satellite image and each cluster center point, and classify each pixel point and its closest cluster center point into one category to complete a clustering; Determine whether the clustering conditions of the pixels before and after the clustering are the same, if they are the same, complete the clustering and go to step S1.4; if they are different, continue to step S1.3;
  • Step S2 applying the self-labeling algorithm based on the combination of support vector machine SVM and extreme learning machine ELM classifier to expand the sample set, and obtain the final expanded sample set EL:
  • step S1 On the basis of image segmentation in step S1, the joint training self-labeling algorithm is applied to expand the training set;
  • the labeled sample set is used as the initial training set to create support vector machine SVM and extreme learning machine ELM classifier, and then the expanded labeled sample set is used as the training set to classify the support vector machine SVM and extreme learning machine ELM
  • the labeling stage a part of unlabeled samples are input into the initially trained SVM and extreme learning machine ELM classifiers, and the confidence of each category is output; CTSLAL algorithm is based on confidence higher than 80%. The samples are labeled, and the training and labeling process is repeated iteratively;
  • Step S3 using the D-ELM algorithm to classify the regional land use and/or agricultural planting structure of the expanded sample set EL:
  • step S1 The satellite image that only contains agricultural areas after being divided in step S1 is used as input, and the final enlarged sample set EL obtained in step S2 is used as a new training set, and the D-ELM classifier is finally applied to the entire satellite containing agricultural areas.
  • the land use and/or planting structure of the image is classified; the result of the classification is an image colored with the classification label.
  • the step S2 includes the following steps:
  • the labeled sample set L is a set of manually labeled pixel points with category labels
  • Unlabeled sample set U that is, a large number of pixel points without labeled categories randomly selected in the original image
  • the support vector machine SVM classifier and the extreme learning machine ELM classifier respectively use the labeled sample set L for initial training and learning;
  • the support vector machine SVM classifier and the extreme learning machine ELM classifier respectively obtain a set of samples from the unlabeled sample set U, mark them, and obtain two annotation sets, and apply the co-training-based evaluator to the two annotations
  • the samples with the same label and the confidence level are higher than 80% are added to the collaborative label set CL, and the enlarged sample set EL will be updated accordingly;
  • step S2.5 through the updated expanded sample set EL, repeat step S2.4, retrain the support vector machine SVM classifier and the extreme learning machine ELM classifier, and then apply the collaborative training evaluator to continuously perform the labeling task;
  • the unmarked sample set is continuously marked through step S2.5 until the number of samples in the enlarged sample set EL no longer increases, thereby obtaining the final enlarged sample set EL.
  • the labeled sample set L includes sunflower, corn, and wheat.
  • the pixels in each image of the classification result have a classification, and different colors represent different classifications.
  • the present disclosure solves the difficulty in acquiring field samples and in the case of sparse samples, reasonably improves the quantity and quality of samples, and ensures the classification accuracy and stability of the classifier.
  • the algorithm of the present disclosure reduces the time complexity and space complexity, so that in the face of large-area high-precision images, ordinary hardware equipment can be used to obtain considerable classification results, so that large-area high-precision images can be obtained.
  • the automatic acquisition of agricultural remote sensing planting structure becomes possible.
  • Fig. 1 is the flow chart of the remote sensing identification method of the disclosed agricultural planting structure
  • FIG. 2 is a partial original satellite image of an embodiment of the present disclosure (for example, google earth historical image is used in this description, other satellite original images are acceptable);
  • Fig. 3 is the image after performing unsupervised algorithm k-means clustering to Fig. 2;
  • FIG. 4 is a classified image according to an embodiment of the disclosure.
  • Fig. 5 is the flow chart of the combined self-labeling algorithm of land use and/or planting structure
  • Figure 6 is a legend for each category of remote sensing images.
  • a remote sensing identification method of an agricultural planting structure of the present disclosure combines image segmentation with a self-labeling algorithm. It includes the following steps:
  • Step S1 using the k-means unsupervised learning algorithm to segment the original satellite image:
  • step S1.2 Calculate the distance between the unclustered pixel points in the original satellite image and each cluster center point, and classify each pixel point and its closest cluster center point into one category to complete a clustering; Determine whether the clustering conditions of the pixels before and after the clustering are the same, if they are the same, complete the clustering and go to step S1.4; if they are different, continue to step S1.3;
  • FIG. 6 it is a legend for each category of remote sensing images, including wasteland, sandy land, water body, residential land, bare land, swamp, construction land and agricultural cultivated land.
  • wasteland, sandy land, water body, residential land, bare land, swamp, and construction land are non-agricultural areas, and agricultural arable land is agricultural area.
  • Step S2 applying the self-labeling algorithm based on the combination of support vector machine SVM and extreme learning machine ELM classifier to expand the sample set, and obtain the final expanded sample set EL:
  • CTSLAL joint training self-labeling algorithm
  • the SVM and ELM classifiers are created using the labeled sample set as the initial training set, and then the SVM and the extreme learning machine ELM classifier are updated with the expanded labeled sample set as the training set.
  • the labeling stage a portion of the unlabeled samples are input into the initially trained SVM and extreme learning machine ELM classifiers, and the confidence of each category is output.
  • the CTSLAL algorithm labels samples based on confidence levels higher than 80%, and the training and labeling process is repeated iteratively.
  • the labeled sample set L is a set of artificially labeled pixel points with category labels.
  • the labeled sample set L of the present disclosure includes, for example, sunflower, corn, wheat, and the like.
  • Unlabeled sample set U that is, a large number of pixel points without labeled categories randomly selected in the original image.
  • the collaborative labeling set CL is a set of samples completed by collaborative labeling and self-labeling
  • the enlarged sample set EL is the union of the artificially labeled samples and the collaborative labeling set CL.
  • the support vector machine SVM classifier and the extreme learning machine ELM classifier respectively use the labeled sample set L for initial training and learning;
  • the support vector machine SVM classifier and the extreme learning machine ELM classifier respectively obtain a set of samples from the unlabeled sample set U, mark them, and obtain two annotation sets, and apply the co-training-based evaluator to the two annotations
  • the samples with the same label and the confidence level are higher than 80% are added to the collaborative label set CL, and the enlarged sample set EL will be updated accordingly.
  • Step S2.4 is repeated through the updated expanded sample set EL to retrain the support vector machine SVM classifier and the extreme learning machine ELM classifier, and then apply the collaborative training evaluator to continuously perform the labeling task.
  • the unmarked sample set is continuously marked through step S2.5 until the number of samples in the enlarged sample set EL no longer increases, thereby obtaining the final enlarged sample set EL.
  • Step S3 using the D-ELM algorithm to classify the regional land use and/or agricultural planting structure of the expanded sample set EL:
  • step S1 The satellite image that only contains agricultural areas after being segmented in step S1 is used as input, and the final mid-scale enlarged sample set EL obtained in step S2 is used as a new training set, and the D-ELM classifier is finally applied to the entire satellite containing agricultural areas. Images are classified by land use and/or planting structure.
  • the result of the classification is an image with the color of the classification label.
  • the output part in Figure 1 is the classification result.
  • the pixels in each image have a classification, and different colors represent different classifications.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Health & Medical Sciences (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Health & Medical Sciences (AREA)
  • Multimedia (AREA)
  • Probability & Statistics with Applications (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)

Abstract

本公开属于农业遥感领域，涉及一种农业种植结构的遥感识别方法，采用机器学习算法提取农作物的种植结构，并且满足多种训练数据类型，自动训练后得到高精度农业种植结构栅格化数据。本发明通过非监督算法k-means将相似像元先聚类成单一类别，而后运用自标记算法对样本集进行扩充，最后利用D-ELM算法将合并后的像元类别进行训练和计算。通过训练后的分类器再将所有区域分类判别，这样大大提高了分类器的效率，避免了使用监督分类算法时间空间复杂度过高而导致大区域种植结构分类不可行的问题，同时D-ELM混合分类器在保证分类精度前提下，满足农业遥感种植结构识别的制图精度。

Description

一种农业种植结构的遥感识别方法
本公开要求于2021年04月07日提交中国专利局、申请号为202110371126.X、发明名称为“一种农业种植结构的遥感识别方法”的中国专利申请的优先权，其全部内容通过引用结合在本公开中。
技术领域
本公开属于农业遥感领域，涉及一种农业种植结构的遥感识别方法，采用机器学习算法提取农作物的种植结构，并且满足多种训练数据类型，自动训练后得到高精度农业种植结构栅格化数据。
背景技术
20世纪以来，遥感能准确地获取农业生产信息，成为指导传统农业向信息科学化农业转变的主要技术方法。如今，利用遥感技术可以实现农作物识别及对应种植面积估算、农作物长势监测与产量估计、农作物叶面指数、叶绿素含量监测与养分诊断、农用地提取与动态监测等。因此，遥感和地理信息技术结合已经成为快速收集和定量分析农业信息、实现科学快速决策的基本手段。其中农作物种植结构识别和种植面积估算是进行农业结构调整的依据，也是研究水文模型、农业决策的重要数据源。从20世纪80年代中期，中国就开始利用气象卫星开展小麦、水稻、玉米等大宗作物的面积监测、长势及产量估计等技术研究。极限学习机(Extreme Learning Machine,ELM)因其学习速度快、泛化能力强而被广泛应用于各种分类任务中。
极限学习机(Extreme Learning Machine,ELM)作为一类针对前馈神经网络设计的机器学习算法，在2004年由Guang-Bin Huang、Qin-Yu Zhu和Chee-Kheong Siew提出，其诞生的动机是为了克服传统神经网络算法学习效率低、参数设定繁琐的问题。
ELM的特点是其学习过程不需要调整隐含层节点参数，输入层至 隐含层的特征映射可以是随机的或人为给定的。由于仅需求解输出权重，ELM在本质上是一个线性参数模式(linear-in-the-parameter model)，其学习过程易于在全局极小值收敛。ELM在学习效率和泛化性能上的优势已经在许多领域得到证实，在与反向传播算法(Back-Propagation,BP)和支持向量机(Support Vector Machine,SVM)的比较中，ELM的学习时间显著缩短，且学习精度相当。ELM最初是为传统的分类和回归类问题而设计的，但在随后的研究中，其应用范围被推广至几乎所有机器学习领域，包括聚类和特征学习(representational learning)等，并出现许多变体和改进算法。ELM保持了前馈神经网络的万能近似定理(universal approximation theorem)适用性，而其学习策略为机器学习领域带来了诸多启发。然而，由于其浅层结构特征，ELM无法有效提取数据深层抽象信息。因此应用深度极限学习机，对解决大数据稀疏样本条件下种植结构的识别应用十分重要。
针对卫星遥感图像现有的分类算法难以在大范围高精度图像中实施、运算时间长且耗费大量的内存。对农业种植结构识别应用而言，单一的分类算法以及分类器很难解决实际情况所面对的计算量大，样本稀疏等问题。
在实际工作中聚类算法、SVM、随机森林、ELM等各种分类器都具有其优点与缺点，因此本公开结合各种算法的优势，取长补短，提高分类的效率以及精度。监督分类算法的优势在于可以根据训练数据样本对像元进行类别判别，根据样本的特征来确定像元的归属。监督分类的主要缺陷是必须在分类前定义样本的性质，分类效率低，而非监督分类是在没有样本的条件下，根据像元相似度自动判别归类，没有人工干预的成分，具有速度较快的优势，但当地物光谱的差异比较微小的时候，分类效果不如监督分类效果好，且无监督分类不能判别像元的归类，既为聚类算法。
发明内容
(一)要解决的技术问题
针对卫星遥感图像，现有的分类算法难以在大范围高精度图像中实施，且单一的分类算法以及分类器的分类效率较低。
(二)技术方案
针对上述技术问题，本公开的目的是提供一种农业种植结构的遥感识别方法，通过非监督算法k-means将相似像元先聚类成单一类别，而后运用自标记算法对样本集进行扩充，最后利用D-ELM算法将这些合并后的像元类别训练和计算。通过训练后的再将所有区域分类判别，这样大大提高了分类器的效率，避免了使用监督分类算法时间空间复杂度过高而导致大区域种植结构分类不可行的问题，同时D-ELM混合分类器在保证分类精度前提下，满足农业遥感种植结构识别的制图精度。
为了实现上述目的，本公开提供了如下技术方案：
一种农业种植结构的遥感识别方法，包括如下步骤：
步骤S1、利用k-means无监督学习算法对原始卫星图像进行分割：
S1.1、将原始卫星图像中的像素点分为K个簇，随机选取K个像素点作为初始的聚类中心点；
S1.2、计算原始卫星图像中的未聚类的像素点与各个聚类中心点之间的距离，将各像素点和与其距离最近的聚类中心点归为一类，完成一次聚类；判断此次聚类前后像素点的聚类情况是否相同，若相同，则完成聚类，进行步骤S1.4；若不同，则继续进行步骤S1.3；
S1.3、每分配一个像素点，重新计算该聚类中各像素点的中心点，并将该中心点设为新的聚类中心点，重复步骤S1.2；
S1.4、应用图例对聚类后的图像中各个簇与原始图像人工比对进行分析；然后将非农业区域覆盖的簇进行组合，最终对组合后的非农业区域进行裁剪，得到只包含农业区域的卫星图像；
步骤S2、应用基于支持向量机SVM和极限学习机ELM分类器结合训练的自标记算法对样本集进行扩充，得到最终扩大样本集EL：
在步骤S1图像分割的基础上，应用联合训练自标记算法来扩充训练集；
在训练过程中，使用标记样本集作为初始训练集，创建支持向量机SVM和极限学习机ELM分类器，然后以扩大后的标记样本集作为训练集，对支持向量机SVM和极限学习机ELM分类器进行更新；在标记阶段，将一部分未标记样本输入到经过初始训练的支持向量机SVM和极限学习机ELM分类器中，输出每个类别的置信度；CTSLAL算法根据置信度高于80％对样本进行标记，训练和标记过程是通过迭代重复进行；
步骤S3、利用D-ELM算法对扩大样本集EL的区域土地利用和/或农业种植结构进行分类：
将步骤S1中被分割后的只包含农业区域的卫星图像作为输入，同时将步骤S2中得到的最终扩大样本集EL作为新的训练集合，最终应用D-ELM分类器对整个包含农业区域的卫星图像的土地利用和/或种植结构进行分类；分类的结果是得到带有分类标签色彩的一张图像。
在使用k-means无监督学习算法进行图像分割时，为得到一个稳定的聚类结果，选择参数为K＝50个簇进行聚类。
所述步骤S2包括如下步骤：
S2.1、以标记样本集L和未标记样本集U作为输入；
其中，标记样本集L，为人工标记带有类别标签的像素点集合；
未标记样本集U，即在原始图像中随机抽取的大量没有标记类别的像素点集合；
S2.2、将标记样本集L和协同标记集CL结合，得到一个扩大样本集EL；协同标记集CL为协同标记自标记完成的样本集合，扩大样本集EL为人工标记样本与协同标记集CL的并集；
S2.3、支持向量机SVM分类器和极限学习机ELM分类器分别用标记样本集L进行初始训练学习；
S2.4、支持向量机SVM分类器和极限学习机ELM分类器分别从 未标记样本集U中取得一组样本，进行标记，得到两个注释集，并应用基于协同训练评估器对两个注释集进行比较，将具有相同标注且置信度均高于80％的样本添加到协同标记集CL中，扩大样本集EL将相应地得到更新；
S2.5、通过更新的扩大样本集EL，重复步骤S2.4，重新训练支持向量机SVM分类器和极限学习机ELM分类器，然后应用协同训练评估器持续地执行标记任务；
S2.6、未标记样本集通过步骤S2.5被持续地标记，直到扩大样本集EL中的样本数量不再增加，从而得到最终扩大样本集EL。
所述标记样本集L包括葵花、玉米、小麦。
分类的结果的每一个图像中的像素点都有一个分类，不同颜色代表不同分类。
(三)有益效果
与现有技术相比，本公开的有益效果在于：
1、本公开解决实地样本获取困难，样本稀疏的情况下，合理提高样本的数量以及质量，保证分类器的分类精度以及稳定性。
2、本公开算法与传统分类器相比，降低了时间复杂度以及空间复杂度，使得在面对大区域高精度的图像时，可以用普通的硬件设备求得可观的分类结果，使得大区域的农业遥感种植结构自动获取成为可能。
附图说明
图1为本公开农业种植结构的遥感识别方法的流程图；
图2为本公开实施例的局部原始卫星图像(本描述中举例使用google earth历史影像，其他卫星原始图像均可)；
图3为对图2进行非监督算法k-means聚类后的图像；
图4为本公开实施例的分类后的图像；
图5为土地利用和(或)种植结构联合自标记算法流程图；
图6为遥感图像各个类别的图例。
具体实施方式
下面结合附图和实施例对本公开进行进一步说明。
如图1所示，本公开的一种农业种植结构的遥感识别方法，该方法将图像分割与自标记算法相结合。包括如下步骤：
步骤S1、利用k-means无监督学习算法对原始卫星图像进行分割：
S1.1、将原始卫星图像中的像素点分为K个簇，随机选取K个像素点作为初始的聚类中心点。
优选地，在本公开中，在使用k-means无监督学习算法进行图像分割时，为得到一个稳定的聚类结果，选择参数为K＝50个簇进行聚类。
S1.2、计算原始卫星图像中的未聚类的像素点与各个聚类中心点之间的距离，将各像素点和与其距离最近的聚类中心点归为一类，完成一次聚类；判断此次聚类前后像素点的聚类情况是否相同，若相同，则完成聚类，进行步骤S1.4；若不同，则继续进行步骤S1.3；
S1.3、每分配一个像素点，重新计算该聚类中各像素点的中心点，并将该中心点设为新的聚类中心点，重复步骤S1.2；
S1.4、应用图例对聚类后的图像中各个簇与原始图像人工比对进行分析。然后将非农业区域覆盖的簇进行组合，最终对组合后的非农业区域进行裁剪，得到只包含农业区域的卫星图像。
如图6所示，为遥感图像各个类别的图例，包括荒地、沙地、水体、居民地、裸地、沼泽、建筑用地和农业耕地。其中，荒地、沙地、水体、居民地、裸地、沼泽、建筑用地为非农业区域，农业耕地为农业区域。
步骤S2、应用基于支持向量机SVM和极限学习机ELM分类器结合训练的自标记算法对样本集进行扩充，得到最终扩大样本集EL：
在步骤S1图像分割的基础上，应用联合训练自标记算法(CTSLAL)来扩充训练集。
在训练过程中，使用标记样本集作为初始训练集，创建SVM和 ELM分类器，然后以扩大后的标记样本集作为训练集，对支持向量机SVM和极限学习机ELM分类器进行更新。在标记阶段，将一部分未标记样本输入到经过初始训练的支持向量机SVM和极限学习机ELM分类器中，输出每个类别的置信度。CTSLAL算法根据置信度高于80％对样本进行标记，训练和标记过程是通过迭代重复进行。
S2.1、以标记样本集L和未标记样本集U作为输入。
其中，标记样本集L，为人工标记带有类别标签的像素点集合。如图5所示，本公开的标记样本集L例如包括葵花、玉米、小麦等。
未标记样本集U，即在原始图像中随机抽取的大量没有标记类别的像素点集合。
S2.2、将标记样本集L和协同标记集CL结合，得到一个扩大样本集EL。如图5中所示，协同标记集CL为协同标记自标记完成的样本集合，扩大样本集EL为人工标记样本与协同标记集CL的并集。
S2.3、支持向量机SVM分类器和极限学习机ELM分类器分别用标记样本集L进行初始训练学习；
S2.4、支持向量机SVM分类器和极限学习机ELM分类器分别从未标记样本集U中取得一组样本，进行标记，得到两个注释集，并应用基于协同训练评估器对两个注释集进行比较，将具有相同标注且置信度均高于80％的样本添加到协同标记集CL中，扩大样本集EL将相应地得到更新。
S2.5、通过更新的扩大样本集EL，重复步骤S2.4，重新训练支持向量机SVM分类器和极限学习机ELM分类器，然后应用协同训练评估器可持续地执行标记任务。
S2.6、未标记样本集通过步骤S2.5被持续地标记，直到扩大样本集EL中的样本数量不再增加，从而得到最终扩大样本集EL。
步骤S3、利用D-ELM算法对扩大样本集EL的区域土地利用和/或农业种植结构进行分类：
将步骤S1中被分割后的只包含农业区域的卫星图像作为输入，同 时将步骤S2得到的最终中扩大样本集EL作为新的训练集合，最终应用D-ELM分类器对整个包含农业区域的卫星图像的土地利用和/或种植结构进行分类。
分类的结果是得到带有分类标签色彩的一张图像。
图1中输出的部分就是分类结果。优选地，每一个图像中的像素点都有一个分类，不同颜色代表不同分类。
工业实用性
本公开提供的农业种植结构的遥感识别方法，通过非监督算法k-means将相似像元先聚类成单一类别，而后运用自标记算法对样本集进行扩充，最后利用D-ELM算法将合并后的像元类别训练和计算，可以解决实地样本获取困难的问题。在样本稀疏的情况下，可以合理提高样本的数量以及质量，保证分类器的分类精度以及稳定性，具有很强的工业实用性。

Claims (5)

  1. 一种农业种植结构的遥感识别方法，其特征在于，该方法包括如下步骤：
    步骤S1、利用k-means无监督学习算法对原始卫星图像进行分割：
    S1.1、将原始卫星图像中的像素点分为K个簇，随机选取K个像素点作为初始的聚类中心点；
    S1.2、计算原始卫星图像中的未聚类的像素点与各个聚类中心点之间的距离，将各像素点和与其距离最近的聚类中心点归为一类，完成一次聚类；判断此次聚类前后像素点的聚类情况是否相同，若相同，则完成聚类，进行步骤S1.4；若不同，则继续进行步骤S1.3；
    S1.3、每分配一个像素点，重新计算该聚类中各像素点的中心点，并将该中心点设为新的聚类中心点，重复步骤S1.2；
    S1.4、应用图例对聚类后的图像中各个簇与原始图像人工比对进行分析；然后将非农业区域覆盖的簇进行组合，最终对组合后的非农业区域进行裁剪，得到只包含农业区域的卫星图像；
    步骤S2、应用基于支持向量机SVM和极限学习机ELM分类器结合训练的自标记算法对样本集进行扩充，得到最终扩大样本集EL：
    在步骤S1图像分割的基础上，应用联合训练自标记算法来扩充训练集；
    在训练过程中，使用标记样本集作为初始训练集，创建支持向量机SVM和极限学习机ELM分类器，然后以扩大后的标记样本集作为训练集，对支持向量机SVM和极限学习机ELM分类器进行更新；在标记阶段，将一部分未标记样本输入到经过初始训练的支持向量机SVM和极限学习机ELM分类器中，输出每个类别的置信度；CTSLAL算法根据置信度高于80％对样本进行标记，训练和标记过程是通过迭代重复进行；
    步骤S3、利用D-ELM算法对扩大样本集EL的区域土地利用和/ 或农业种植结构进行分类：
    将步骤S1中被分割后的只包含农业区域的卫星图像作为输入，同时将步骤S2中得到的最终扩大样本集EL作为新的训练集合，最终应用D-ELM分类器对整个包含农业区域的卫星图像的土地利用和/或种植结构进行分类；分类的结果是得到带有分类标签色彩的一张图像。
  2. 根据权利要求1所述的农业种植结构的遥感识别方法，其特征在于，在使用k-means无监督学习算法进行图像分割时，为得到一个稳定的聚类结果，选择参数为K＝50个簇进行聚类。
  3. 根据权利要求1所述的农业种植结构的遥感识别方法，其特征在于，所述步骤S2包括如下步骤：
    S2.1、以标记样本集L和未标记样本集U作为输入；
    其中，标记样本集L，为人工标记带有类别标签的像素点集合；
    未标记样本集U，即在原始图像中随机抽取的大量没有标记类别的像素点集合；
    S2.2、将标记样本集L和协同标记集CL结合，得到一个扩大样本集EL；协同标记集CL为协同标记自标记完成的样本集合，扩大样本集EL为人工标记样本与协同标记集CL的并集；
    S2.3、支持向量机SVM分类器和极限学习机ELM分类器分别用标记样本集L进行初始训练学习；
    S2.4、支持向量机SVM分类器和极限学习机ELM分类器分别从未标记样本集U中取得一组样本，进行标记，得到两个注释集，并应用基于协同训练评估器对两个注释集进行比较，将具有相同标注且置信度均高于80％的样本添加到协同标记集CL中，扩大样本集EL将相应地得到更新；
    S2.5、通过更新的扩大样本集EL，重复步骤S2.4，重新训练支持向量机SVM分类器和极限学习机ELM分类器，然后应用协同训练评估器持续地执行标记任务；
    S2.6、未标记样本集通过步骤S2.5被持续地标记，直到扩大样本 集EL中的样本数量不再增加，从而得到最终扩大样本集EL。
  4. 根据权利要求3所述的农业种植结构的遥感识别方法，其特征在于，所述标记样本集L包括葵花、玉米、小麦。
  5. 根据权利要求1所述的农业种植结构的遥感识别方法，其特征在于，分类的结果的每一个图像中的像素点都有一个分类，不同颜色代表不同分类。
PCT/CN2022/085640 2021-04-07 2022-04-07 一种农业种植结构的遥感识别方法 WO2022214039A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110371126.XA CN112906666A (zh) 2021-04-07 2021-04-07 一种农业种植结构的遥感识别方法
CN202110371126.X 2021-04-07

Publications (1)

Publication Number Publication Date
WO2022214039A1 true WO2022214039A1 (zh) 2022-10-13

Family

ID=76110037

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/085640 WO2022214039A1 (zh) 2021-04-07 2022-04-07 一种农业种植结构的遥感识别方法

Country Status (2)

Country Link
CN (1) CN112906666A (zh)
WO (1) WO2022214039A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116091938A (zh) * 2022-12-29 2023-05-09 宁波大学 一种针对单季稻种植面积的多源遥感监测方法
CN116453003A (zh) * 2023-06-14 2023-07-18 之江实验室 一种基于无人机监测智能识别水稻生长势的方法和系统
CN117871792A (zh) * 2024-03-13 2024-04-12 河北省建筑科学研究院有限公司 用于对园区绿植碳汇的动态监测方法及系统

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112906666A (zh) * 2021-04-07 2021-06-04 中国农业大学 一种农业种植结构的遥感识别方法
CN114612794B (zh) * 2022-03-01 2024-06-07 中国农业大学 一种细碎化农业区土地覆被与种植结构的遥感识别方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110235900A1 (en) * 2010-03-26 2011-09-29 Fatih Porikli Method for Training Multi-Class Classifiers with Active Selection and Binary Feedback
CN105279519A (zh) * 2015-09-24 2016-01-27 四川航天系统工程研究所 基于协同训练半监督学习的遥感影像水体提取方法及系统
CN105303198A (zh) * 2015-11-17 2016-02-03 福州大学 一种自定步长学习的遥感影像半监督分类方法
CN108960276A (zh) * 2018-05-08 2018-12-07 南京理工大学 提升光谱图像监督分类性能的样本扩充与一致性判别方法
CN109146876A (zh) * 2018-09-14 2019-01-04 四川省安全科学技术研究院 一种基于高分遥感影像的矿山环境变化检测方法
CN112906666A (zh) * 2021-04-07 2021-06-04 中国农业大学 一种农业种植结构的遥感识别方法

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101373519B (zh) * 2007-08-20 2011-01-19 富士通株式会社 字符识别装置和方法
CN101720654A (zh) * 2009-12-08 2010-06-09 中国农业大学 草坪节水灌溉预报方法及系统
US20130097103A1 (en) * 2011-10-14 2013-04-18 International Business Machines Corporation Techniques for Generating Balanced and Class-Independent Training Data From Unlabeled Data Set
CN104992184B (zh) * 2015-07-02 2018-03-09 东南大学 一种基于半监督极限学习机的多类图像分类方法
CN105550744A (zh) * 2015-12-06 2016-05-04 北京工业大学 一种基于迭代的神经网络聚类方法
CN109447151A (zh) * 2018-10-26 2019-03-08 成都国星宇航科技有限公司 一种基于深度学习的遥感数据分析方法
CN109460795A (zh) * 2018-12-17 2019-03-12 北京三快在线科技有限公司 分类器训练方法、装置、电子设备及计算机可读介质
CN109784192B (zh) * 2018-12-20 2023-04-18 西安电子科技大学 基于超像素特征提取神经网络算法的高光谱影像分类方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110235900A1 (en) * 2010-03-26 2011-09-29 Fatih Porikli Method for Training Multi-Class Classifiers with Active Selection and Binary Feedback
CN105279519A (zh) * 2015-09-24 2016-01-27 四川航天系统工程研究所 基于协同训练半监督学习的遥感影像水体提取方法及系统
CN105303198A (zh) * 2015-11-17 2016-02-03 福州大学 一种自定步长学习的遥感影像半监督分类方法
CN108960276A (zh) * 2018-05-08 2018-12-07 南京理工大学 提升光谱图像监督分类性能的样本扩充与一致性判别方法
CN109146876A (zh) * 2018-09-14 2019-01-04 四川省安全科学技术研究院 一种基于高分遥感影像的矿山环境变化检测方法
CN112906666A (zh) * 2021-04-07 2021-06-04 中国农业大学 一种农业种植结构的遥感识别方法

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116091938A (zh) * 2022-12-29 2023-05-09 宁波大学 一种针对单季稻种植面积的多源遥感监测方法
CN116091938B (zh) * 2022-12-29 2023-08-18 宁波大学 一种针对单季稻种植面积的多源遥感监测方法
CN116453003A (zh) * 2023-06-14 2023-07-18 之江实验室 一种基于无人机监测智能识别水稻生长势的方法和系统
CN116453003B (zh) * 2023-06-14 2023-09-01 之江实验室 一种基于无人机监测智能识别水稻生长势的方法和系统
CN117871792A (zh) * 2024-03-13 2024-04-12 河北省建筑科学研究院有限公司 用于对园区绿植碳汇的动态监测方法及系统
CN117871792B (zh) * 2024-03-13 2024-05-14 河北省建筑科学研究院有限公司 用于对园区绿植碳汇的动态监测方法及系统

Also Published As

Publication number Publication date
CN112906666A (zh) 2021-06-04

Similar Documents

Publication Publication Date Title
WO2022214039A1 (zh) 一种农业种植结构的遥感识别方法
Lu et al. Cultivated land information extraction in UAV imagery based on deep convolutional neural network and transfer learning
Wang et al. YOLOv3‐Litchi Detection Method of Densely Distributed Litchi in Large Vision Scenes
CN104408466B (zh) 基于局部流形学习构图的高光谱遥感图像半监督分类方法
CN110598564B (zh) 基于OpenStreetMap的高空间分辨率遥感影像迁移学习分类方法
CN109448127A (zh) 一种基于无人机遥感的农田高精度导航地图生成方法
CN112613428B (zh) 基于平衡损失的Resnet-3D卷积牛视频目标检测方法
Chen et al. Agricultural remote sensing image cultivated land extraction technology based on deep learning
CN112329559A (zh) 一种基于深度卷积神经网络的宅基地目标检测方法
WO2020232942A1 (zh) 基于农田图像的卷积神经网络构建方法及其系统
CN111723711A (zh) 基于Pléiades和面向对象的地膜信息提取方法及系统
CN104573701B (zh) 一种玉米雄穗性状的自动检测方法
Xia et al. Phenology-based decision tree classification of rice-crayfish fields from Sentinel-2 imagery in Qianjiang, China
Xin et al. A data augmented method for plant disease leaf image recognition based on enhanced GAN model network
Gao et al. Maize seedling information extraction from UAV images based on semi-automatic sample generation and Mask R-CNN model
CN115063655A (zh) 一种融合超列的类激活映射图生成方法
Wang et al. Multimodal Poisson gamma belief network
Su Superpixel-based principal component analysis for high resolution remote sensing image classification
Fan et al. EDNET: A mesoscale eddy detection network with multi-modal data
Liu et al. Integration transformer for ground-based cloud image segmentation
Nagila et al. The effectiveness of machine learning and image processing in detecting plant leaf disease
CN110210376A (zh) 一种基于深度学习的高分遥感影像西北地区农膜提取方法
CN116563205A (zh) 基于小目标检测和改进YOLOv5的麦穗计数检测方法
Wei et al. Application of remote sensing technology in crop estimation
CN113643248B (zh) 一种基于改进卷积神经网络的小麦生育进程监测方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22784105

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 22784105

Country of ref document: EP

Kind code of ref document: A1