| 摘要: |
| 【目的】应用浅层结构的机器学习分类器和高空间分辨率影像实现休耕区绿肥、粮食
及经济作物快速准确分类。【方法】利用分辨率为5 m 的RapidEye 影像，以云南省石林县部
分休耕试点区为研究区，使用Softmax 浅层机器学习分类器对研究区内绿肥作物、水稻、玉
米及烟草等4 种典型作物进行遥感识别与空间信息提取，并以极大似然分类法为参照，通过
地面样方数据验证该方法的精度。【结果】基于Softmax 方法的4 种典型作物分类的总体精
度和Kappa 系数分别为85.98% 和0.815 7，比极大似然分类高4.59% 和0.061 7；绿肥、水
稻、烟草的生产者精度和用户精度均达到84% 以上，玉米则低于75%，原因是绿肥、水
稻、烟草3 种作物种植较为集中，而玉米种植地块面积小且极为分散；绿肥与烟草错分问题
较明显，影响因素为“同物异谱、异物同谱”。【结论】基于Softmax 的浅层机器学习分类器
提高了分类精度，文章研究结果可为使用浅层机器学习方法快速准确掌握休耕情况提供参
考。 |
| 关键词: 遥感 Softmax 分类器 休耕 作物分类 |
| DOI：10.12105/j.issn.1672-0423.20190202 |
| 分类号: |
| 基金项目:四川省科技厅软科学研究项目“基于高分六号遥感影像的四川粮食作物布局研究”（2019JDR0121）；四
川省科技厅应用基础研究项目“基于空间大数据的乡村地区土地利用变化研究”（2019YJ0608）；四川省
重点研发项目“基于物联网+ 遥感技术的智慧农业研究”（2017GZ0160）；四川省省院省校合作项目“基
于大数据机器学习与冠层反射率模型结合的水稻叶面积指数提取技术”（2018JZ0054）；四川省应用基础
研究项目“基于互联网+ 多阶段遥感反演的区域水稻参数逐田块监测技术研究”（2017JY0284）；四川
省财政创新能力提升工程青年基金“基于冠层反射率模型多阶段反演的逐地块水稻参数采集技术研究”
（2017QNJJ-023） |
|
| Shallow learning classification of summer cropsin rocky desertified fallow pilot area |
|
Dong Xiuchun, Jiang Yi, Huang Ping, Li Zongnan※, Liu Ke
|
|
Institute of Remote Sensing Application，Sichuan Academy of Agricultural Sciences，Sichuan Chengdu 610066，China
|
| Abstract: |
| ［Purpose］Fallow monitoring is an important part of the fallow pilot project under supervision of the Ministry of Agriculture and Rural Affair. Remote sensing is able to monitor the land whether is leaving fallow or planting green manure for protecting the land capability. The monitoring result is of great significance for decision makers to know the situation of fallow in the regional scale. To achieve the fast and accurate classification of multiple crops，including green manure，grain and cash crops in the fallow region，the machine learning classifier with shallow structure and high spatial resolution remote sensing image is applied. The aim of this study is to provide reference for monitoring the land fallow situation in the fallow region with the method of the shallow learning.［ Method］In this study，part of the fallow pilot region in Shilin county was selected as the study area. The Rapideye image with the spatial resolution of 5m was used，with the support of Softmax classifier which belongs to the machine learning of shallow structure. The recognition and spatial information extraction of four typical crops，such as green manure，rice，corn and tobacco were carried out in the study area. By using the data of insitu survey，the accuracy was evaluated and compared with the results of maximum likelihood classification.［ Result］The results revealed that，the overall accuracy and Kappa coefficients of the classification based on the Softmax classifier were 85.98% and 0.815 7，respectively，slightly higher than the maximum likelihood method which has the results of 81.39% and 0.754 0. The producer precision and user precision of green manure，rice and tobacco were higher than 84%，while corn was less than 75%. The reason was that the planting of green manure，rice and tobacco are relatively concentrated，while the planting of corn was scattered. There were obvious classification errors，because of the similarity between green manure and tobacco. ［Conclusion］The results indicate that Softmax classifier can improve the accuracy of multiple crops classification in fallow region. This method can provide reference for the application of shallow machine learning in fallow region. |
| Key words: remote sensing softmax classifier fallow crop classification |
