| 摘要: |
| 目的 农作物精细分类对于农作物长势监测、产量预估、灾害评估、保障国家粮食安全具有重要意义。高光谱遥感影像具有丰富的光谱波段，能够探测到各类农作物之间细微差别，已逐渐成为分类的理想数据源。方法 研究以由AVIRIS传感器收集的美国加利福尼亚州南部萨利纳斯山谷的农作物区域的高光谱数据为数据源，提出了一种基于条件随机场的高光谱遥感农作物精细分类方法，利用SVM分类器计算各类地物的概率，并定义为条件随机场的一元势函数以融合空间特征信息；将空间平滑项和局部类别标签成本项加入到二元势函数中，以考虑空间背景信息，并保留各类别中的详细信息。最后与传统的最小距离法和SVM算法进行比较。结果 文章提出的方法较最小距离分类法、SVM传统方法相比，整体精度分别提高了16%和2%，除了C15类（葡萄园3）精度为72.32%与74.11%外，各类地物精度均在94%以上，各种“椒盐”噪声与分类混淆现象得到了改善。结论 实验结果表明，该方法在农作物精细分类应用中具有较大优势。 |
| 关键词: 高光谱遥感 农作物精细分类 条件随机场 |
| DOI：10.12105/j.issn.1672-0423.20180307 |
| 分类号: |
| 基金项目:国家重点研发计划课题（2017YFB0504202）、国家自然科学基金优秀青年科学基金项目（41622107）、湖北省技术创新专项重大项目（2018ABA078）、空间数据挖掘与信息共享教育部重点实验室开放基金（2018LSDMIS05）、农业部农业遥感重点实验室开放基金20170007国家重点研发计划课题（2017YFB0504202）、国家自然科学基金优秀青年科学基金项目（41622107）、湖北省技术创新专项重大项目（2018ABA078）、空间数据挖掘与信息共享教育部重点实验室开放基金（2018LSDMIS05）、农业部农业遥感重点实验室开放基金（20170007） |
|
| Hyperspectral remote sensing image crop fine classification based on conditional random field |
|
Yu Ming1,Wei Lifei1,Yin Feng2,Li Dandan3,Huang Qingbin4
|
|
1.Faculty of Resources and Environmental Science，Hubei University，Wuhan 430062，China;2.Hubei Provincial Institute of Land and Resources，Wuhan 430062，China;3.Institute of Agricultural Resources and Regional Planning ，Chinese Academy of Agricultural Sciences / Key Laboratory of Agricultural Remote Sensing，Ministry of Agriculture，Beijing 100081，China;4.Shenzhen Cadastral Surveying and Mapping Brigade，Shenzhen 518000，China
|
| Abstract: |
| Purpose Fine classification of crops is of great significance for crop growth monitoring，yield estimation，and disaster assessment.Hyperspectral remote sensing images have a rich spectral band that can detect subtle differences between various crops，and it has been gradually becoming an ideal data source for classification.Methods Based on the hyperspectral agricultural data of the Salinas Valley in southern California，America，this paper proposes a crop fine classification method based on conditional random field hyperspectral remote sensing images，which uses SVM classifier to calculate the probability of various types of features，and is defined as a unary potential function of the conditional random field to fuse spatial feature information；the spatial smoothing term and the local category label cost term are added to the pairwise potential function to consider the spatial background information and retain the detailed information in each category.Finally，it is compared with the traditional minimum distance method and SVM algorithm.Results Compared with the traditional method，the proposed method improves the overall accuracy by 16% and 2% respectively.Except for the C15 （Vineyard 3）class’s accuracy of 72.32% and 74.11%，the accuracy of all kinds of crops is above 94%.Various “salt and salt” noises and the confusion with classification have been improved.Conclusion The experimental results show that the method has great advantages in the fine classification of crops. |
| Key words: hyperspectral remote sensing fine classification of crops conditional random field |
