地球信息科学学报 ›› 2021, Vol. 23 ›› Issue (6): 1071-1081.doi: 10.12082/dqxxkx.2021.200546

• 遥感科学与应用技术 • 上一篇    下一篇

特征优选与卷积神经网络在农作物精细分类中的应用研究

刘戈1,4(), 姜小光1,3, 唐伯惠2,4,*()   

  1. 1.中国科学院大学,北京 100049
    2.昆明理工大学国土资源工程学院,昆明 650093
    3.中国科学院空天信息创新研究院 定量遥感信息技术重点实验室,北京 100094
    4.中国科学院地理科学与资源研究所 资源与环境信息系统国家重点实验室,北京 100101
  • 收稿日期:2020-09-21 修回日期:2021-01-04 出版日期:2021-06-25 发布日期:2021-08-25
  • 通讯作者: 唐伯惠
  • 作者简介:刘 戈(1994— ),女,北京人,硕士生,主要从事定量遥感研究。E-mail: liuge18@mails.ucas.edu.cn
  • 基金资助:
    中国科学院大科学项目(181811KYSB20160040);中国科学院先导项目(XDA13030402)

Application of Feature Optimization and Convolutional Neural Network in Crop Classification

LIU Ge1,4(), JIANG Xiaoguang1,3, TANG Bohui2,4,*()   

  1. 1. University of Chinese Academy of Sciences, Beijing 100049, China
    2. Faculty of Land and Resource Engineering, Kunming University of Science and Technology, Kunming 650093, China
    3. Key Laboratory of Quantitative Remote Sensing Information Technology, Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100094, China
    4. State Key Laboratory of Resources and Environmental Information System, Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing 100101, China
  • Received:2020-09-21 Revised:2021-01-04 Online:2021-06-25 Published:2021-08-25
  • Contact: TANG Bohui
  • Supported by:
    Bureau of International Cooperation Chinese Academy of Science(181811KYSB20160040);Strategic Priority Research Program of The Chinese Academy of Sciences(XDA13030402)

摘要:

农作物的精细分类一直是农业遥感领域的热点,对农作物估产和种植结构监管有重要意义。深度学习的出现为农作物分类准确性的提升提供了新的思路。本文提出一种特征优选与卷积神经网络(Convolutional Neural Networks, CNN)相结合的多光谱遥感农作物分类方法,用以解决精细分类问题。实验以哨兵2号遥感影像为数据源,基于多光谱遥感影像的波段反射率与包括归一化植被指数在内的10种植被指数,利用Relief F算法进行特征增强与优选,获取最优特征集,从而设计出基于特征优选的CNN分类方法,并对河南省原阳县主要农作物水稻、玉米、花生进行分类识别与制图,分类精度达到96.39%。同时,选用支持向量机、CNN方法分别对研究区农作物进行分类识别。对比分析3种方法的分类结果,发现本文提出的基于最优特征集的CNN农作物分类方法表现最优,CNN方法次之,支持向量机方法表现最差。实验结果表明:① 利用Relief F算法能够对特征贡献度进行排序,完成特征筛选,得到包含24个特征的最优特征子集,训练精度达到99.89%;② 基于最优特征集的CNN方法能够在最大程度上提取高精度差异性特征,实现对农作物的精细分类,且相比CNN和支持向量机的农作物分类方法,本文方法表现更佳。

关键词: 农作物分类, 遥感, CNN, 深度学习, Relief F, 特征优选, 植被指数, 多光谱

Abstract:

Fine-scale crop classification has always been a hot topic in the field of agricultural remote sensing, which is of great significance for crop yield estimation and planting structure supervision. The emergence of deep learning provides a new way to improve the accuracy of crop classification. Recently, the Convolutional Neural Network (CNN), a representative algorithm of deep learning, shows obvious advantages in processing high-dimensional remote sensing data. However, the application of CNN in crop classification based on multispectral data is still rare, and the classification accuracy dependent on the different feature information of crops is hard to evaluate. In this paper, a crop classification method based on feature selection and CNN for multispectral remote sensing data is proposed to improve fine crop classification. This study used Sentinel-2 remote sensing images as data source. Based on the reflectance of 13 multispectral bands and 10 vegetation indices including normalized difference vegetation index, ratio vegetation index, enhanced vegetation index, etc., the Relief F algorithm was used to rank the contribution of multidimensional features. According to the rank of feature contribution, the features with high contribution were selected and optimized by group training to obtain the best feature collection. Therefore, a CNN-based classification method based on feature selection was designed. Based on this, we classified and mapped the main crops including rice, corn, and peanut in Yuanyang County, Henan Province, with an overall classification accuracy of 96.39%. Meanwhile, the support vector machine and simple CNN were also used to classify the main crops in the research area for comparison. We found that the CNN-based classification method based on the optimal feature collection had the highest classification accuracy, followed by simple CNN, and the support vector machine had the worst performance. The main conclusions of this research are as follows: (1) The Relief F algorithm was effective to sort the contribution of different features. In total, we obtained 24 optimal feature subsets, with a training accuracy of 99.89%; (2) The CNN-based classification method using the optimal feature collection can extract the high-precision difference in features to the greatest extent and realize the fine-scale classification of crops. Compared with simple CNN and support vector machine, the CNN method based on the optimal feature collection has obvious advantages.

Key words: crop classification, remote sensing, CNN, deep learning, Relief F, feature selection, vegetation index, multi-spectrum