地球信息科学学报 ›› 2019, Vol. 21 ›› Issue (6): 918-927.doi: 10.12082/dqxxkx.2019.180424

• 遥感科学与应用技术 • 上一篇    下一篇

基于多层神经网络与Sentinel-2数据的大豆种植区识别方法

田富有1,2(), 吴炳方1,2,*(), 曾红伟1, 何昭欣1,2, 张淼1, JoséBofana1,2   

  1. 1. 中国科学院遥感与数字地球研究所 遥感科学国家重点实验室,北京 100101
    2. 中国科学院大学,北京 100049
  • 收稿日期:2018-08-30 修回日期:2018-12-27 出版日期:2019-06-15 发布日期:2019-07-03
  • 通讯作者: 吴炳方 E-mail:tianfy@radi.ac.cn;wubf@radi.ac.cn
  • 作者简介:

    作者简介:田富有(1994-),男,山东临沂人,博士生,主要从事基于大数据技术的可持续发展研究。E-mail: tianfy@radi.ac.cn

  • 基金资助:
    中国科学院科技服务网络计划(STS计划)项目(KFJ-STS-ZDTP-009);国家自然科学基金项目(41561144013、41861144019、41701496)

Identifying Soybean Cropped Area with Sentinel-2 Data and Multi-Layer Neural Network

Fuyou TIAN1,2(), Bingfang WU1,2,*(), Hongwei ZENG1, Zhaoxin HE1,2, Miao ZHANG1, Bofana José1,2   

  1. 1. State Key Laboratory of Remote Sensing Science , Institute of Remote Sensing and Digital Earth, Chinese Academy of Science, Beijing 100101, China
    2. College of Resources and Environment, University of Chinese Academy of Sciences, Beijing 100049, China
  • Received:2018-08-30 Revised:2018-12-27 Online:2019-06-15 Published:2019-07-03
  • Contact: Bingfang WU E-mail:tianfy@radi.ac.cn;wubf@radi.ac.cn
  • Supported by:
    Science and Technology Service Network Initiative, No.KFJ-STS-ZDTP-009;the National Natural Science Foundation of China, No.41561144013, 41861144019, 41701496

摘要:

大豆作为全球最重要的油料作物,是中国进口的大宗农产品,对其种植区的精准识别是决策制定、种植结构调整基础,对国家粮食安全有重要意义。本文利用Sentinel-2作为数据源,利用多层神经网络方法与对大豆进行提取,并与随机森林、决策树、支持向量机等机器学习进行对比,发现F1-Socre指标显示多层神经网络的分类精度最高,为93.53%,其次为随机森林、支持向量机、决策树。将神经网络分类结果与SLIC面向对象分割聚合之后,结果既忽略了同一地块的微小差别,又区分出了不同地块的作物差异,很好的体现了大豆的分布。Sentinel-2数据是进行大尺度大豆种植监测的绝佳数据源,大豆与玉米等其他作物在第二个红边波段的反射率有较为明显的差异。多层神经网络方法在图像分类任务中表现出色,结合图像分割算法精度可达到95.51%,可以满足大豆种植面积监测的需求。

关键词: 大豆提取, 多层神经网络, SLIC分割, Sentinel-2数据, 红边波段

Abstract:

As the most important oil crop in the world, soybean is a large-scale agricultural product that China imports. The accurate identification of its planting area is the basis for decision-making and planting structure adjustment, and is of great significance to national food security. Sentinel-2 was used as a data source and multi-layer neural network was employed to map soybean cropped area. Besides, visible and infrared bands, three red-edge bands were also selected after radiation and atmospheric correction using the Sentinel-2 Toolbox. According to our test, 8-hidden-layer conducted using Scikit-learn package in Python2.7 was the optimal structure for identifying soybean and other crops. Simple linear iterative clustering (SLIC), the state-of-art segmentation algorithm, was performed to segment the remote sensed image. This method combined five-dimensional color and the image plane space to efficiently generate compact and nearly uniform super pixels. To remove the “salt and pepper effect”, the pixel-based result was integrated with the object output from the SLIC. If the pixel as soybean in an object accounted for less than 50%, this object was eliminated in the fusion map. The results showed that the overall accuracy of multi-layer neural network was 93.95%, which was highest and followed by the support vector machine, decision tree, and random forest. Then, the neural network classification was selected as the best result to integrate with SLIC object-oriented segmentation, and the results ignored the small differences of the same land and distinguish the crop differences of different blocks compared with the segmentation in eCognition software. Sentinel-2 data is an appropriate data source for large-scale soybean planting mapping. According to feature importance derived from the random forest classifier, near-infrared band is the most critical feature for classification, followed by third red edge band (Band 7), fourth red edge 4 band (Band 8), red band, and second red edge band (Band 6). The reflectance values of soybeans and other crops in the second red edge band were different, indicating a huge potential in crop type identifying. In the future, the red edge band can be introduced more into crop type even landscape classification. The multi-layer neural network method performs well in the image classification task and had similar or better overall accuracy value compared with other outstanding machine learning classifier including SVM, decision tree, and random forest. Combined with the image segmentation algorithm, such as SLIC, multi-layer neural network can map soybean cropped area with an accuracy high up to 95.51%, which can serve for soybean planting area monitoring in a large area.

Key words: Soybean mapping, multi-layer neural network, SLIC segmentation, Sentinel-2, red edge band.