基于多时相遥感数据的农作物分类研究

发布时间：2018-03-01 13:00

本文关键词： 农作物分布遥感多源数据短时间序列参考NDVI　出处：《中国科学院大学(中国科学院遥感与数字地球研究所)》2017年博士论文　论文类型：学位论文

【摘要】：快速获取与分析农作物的种植分布是农作物长势监测的基础。遥感技术以其快速、无损和客观等特点,广泛应用于农作物种植面积提取。本文针对目前大面积农作物遥感提取工作中存在的30m影像时间分辨率不足、地面样本不足和农作物识别时效性较低的实际问题,以我国新疆和美国堪萨斯州为实验区,从提高遥感数据的时间分辨率、使用时间序列长度较短的数据提前获取农作物种植分布结果和减小农作物遥感提取对训练样本数量的需求三个方面展开分析,为大区域农作物遥感提取提供了新的方案。基于现有研究区和数据,得出如下结论:1.评估Landsat TM数据与HJ数据多光谱波段的相似性,发现TM数据和HJ数据拥有类似的多光谱波段。在此基础上,结合两种数据源获得空间分辨率为30m,时间分辨率为15天的影像时间序列,并使用合成的数据进行农作物提取。结果表明,一部分优选的时相可以实现高精度农作物分类,而更高的时间分辨率增加了在最优时相获得影像的可能性,是进行大面积30m农作物分布制图的数据基础。2.以我国新疆和美国堪萨斯州为实验区,分析时间序列长度对农作物识别精度的影响,探讨使用时间序列长度较短的遥感数据进行农作物早期提取的可行性。新疆喀什实验区的结果表明4~6月影像具备识别作物的潜力,研究区主要农作物相互之间JM距离高于1.6,分类总体精度为85.16%。新疆博乐的实验结果表明,识别棉花、葡萄、玉米和打瓜需要的时间序列长度分别为160天、100天、160天和130天,作物生长期和NDVI峰值期(第60~80天和第100天)是识别夏季作物的最佳时相。美国堪萨斯州的实验结果表明,主要农作物的可分性、分类精度和分类置信度在时间序列长度为5个月(4~8月)时达到饱和,且NDVI对分类的贡献最大。3.提出M-Voting和P-Fusion两种融合多种分类器结果的方法,并比较融合分类器方法和单分类器法对训练样本数量的需求。实验结果表明M-Voting法和P-Fusion法在分类的训练样本较少时(少于500),分类精度高于单分类器。当训练样本数量较大时(大于4000),单分类器的分类精度较高,融合分类器不能进一步提高分类精度。另外,基于象元和基于对象的分类方法获得的分类精度类似,但基于对象方法获得的分类结果“椒盐噪声”较少,更接近地表农作物的真实分布。4.提出不能获得地面调查样本时,使用历史数据建立各种作物的参考NDVI时间序列曲线,并使用参考曲线进行30m农作物提取的方法。该方法在博乐实验区和玛纳斯实验区获得的总体分类精度分别为87.13%和83.38%。由于该方法使用MODIS数据获得参考时间序列曲线,因而在分类时需要将参考曲线的MODIS NDVI转换为Landsat NDVI或HJ NDVI,从而引入误差。为避免NDVI转换引起的误差,进一步改进了基于参考曲线进行作物识别的方法。使用2006~2013年堪萨斯州的历史样本直接获得分类年(2014年)的训练样本,并使用这些训练样本进行了2014年30m农作物提取。在获得的5412个训练样本中,5259个样本的作物类别与美国农业部提供的Crop Data Layer(CDL)数据相同。使用这些训练样本提取作物时,农作物提取精度高于90%。
[Abstract]:Fast acquisition and analysis of planting distribution of crops is the basis of crop condition monitoring. Remote sensing technology with its rapid, nondestructive and objective characteristics, widely used in crop planting area extraction based on 30m images. The existing large area remote sensing extraction in time resolution, the actual problem of insufficient sample and ground crop identification timeliness low, in China's Xinjiang and Kansas into the experimentation area, from improving the time resolution of remote sensing data, using short time series data to obtain crop distribution results and three aspects of Remote Sensing Extraction decreases the number of training samples on demand analysis, which provides a new solution for large area remote sensing the existing research area and data extraction. Based on the conclusions are as follows: 1. evaluation of Landsat TM data and HJ data of multi spectral band phase Iconicity, TM data and HJ data with multi spectral bands similar. On this basis, combined with the two kinds of data sources to obtain spatial resolution of 30m, the time resolution of image time series for 15 days, and the crops from the use of synthetic data. The results show that it can achieve high precision crop classification are a part of selection when the higher time resolution increases the possibility of the optimal phase image is obtained, data based.2. 30m crop distribution drawing large area in China's Xinjiang and Kansas into the experimentation area, analysis of time series effects of length on the crop identification accuracy, to evaluate the feasibility of crop early extraction of remote sensing data the use of short time series. The results show that the 4~6 Xinjiang Kashi experimentation area shadow like identify crop potential of main crops in study area between JM distance More than 1.6, the overall classification accuracy of 85.16%. Xinjiang bole. The experimental results show that the identification of cotton, grapes, maize and melon to the length of time series were 160 days, 100 days, 160 days and 130 days, crop growth period and NDVI peak period (60~80 days and 100th days) is the best identification of summer crops when the United States of Kansas. The experimental results show that the main crops of separability, classification accuracy and confidence in the length of time series was 5 months (4~8 months) reached saturation, and the contribution of NDVI to the classification of the maximum.3. M-Voting and P-Fusion two methods of integration of multiple classifier results, and compare the fusion classifier method and a single classifier method demand for the number of training samples. The experimental results show that M-Voting method and P-Fusion method in the classification of the training sample is small (less than 500), higher accuracy than a single classifier. When a large number of training samples (more than 4000), high classification accuracy of single classifier, classifier fusion can further improve the classification accuracy. In addition, based on the image element and classification accuracy classification method based on object is similar, but the classification results obtained based on the object of salt and pepper noise "less is more close to the real distribution of.4. surface of crops can not get ground investigation samples, using historical data to establish reference NDVI time series curve of various crops, and 30m crops are extracted using the reference curve. The overall classification accuracy obtained in the experimentation area and experimentation area of Manasi Bole were 87.13% and 83.38%. due to the method of using MODIS data to obtain reference time series curve, so when the classification needs to be converted the MODIS NDVI Landsat reference curve of NDVI or HJ NDVI, thus introducing error. In order to avoid the error caused by the conversion of NDVI into. A improved method of crop identification based on the reference curve. The use of 2006~2013 in Kansas state history samples directly obtained classification year (2014) of the training samples, and the 2014 30m crops are extracted using these training samples. In 5412 samples obtained in 5259 samples of the crop category and the U.S. Department of agriculture Crop Data Layer (CDL) the same data. Using these training samples to extract crop, crop extraction accuracy is higher than that of 90%.

【学位授予单位】：中国科学院大学(中国科学院遥感与数字地球研究所)
【学位级别】：博士
【学位授予年份】：2017
【分类号】：S127

【相似文献】