浙江农业学报 ›› 2022, Vol. 34 ›› Issue (12): 2767-2777.DOI: 10.3969/j.issn.1004-1524.2022.12.20

• 生物系统工程 • 上一篇    下一篇

基于Sentinel-2多时相影像的果树种植区遥感提取

周欣兴1(), 赵林1,*(), 张文杰1, 谭昌伟2, 李刚波1, 石梦云1, 张婷1, 杨峰1   

  1. 1.江苏徐淮地区徐州农业科学研究所,江苏 徐州 221121
    2.扬州大学 农学院,江苏 扬州 225009
  • 收稿日期:2021-10-18 出版日期:2022-12-25 发布日期:2022-12-26
  • 通讯作者: 赵林
  • 作者简介:*赵林,E-mail: zhao_20020347@163.com
    周欣兴(1996—),男,安徽铜陵人,硕士,研究实习员,主要从事农业遥感监测研究。E-mail: 18355090421@163.com
  • 基金资助:
    国家重点研发计划(2018YFD0300805);国家自然科学基金(32071902);徐州市科技计划(KC18128);江苏省政策引导类计划(苏北科技专项)(SZ-XZ2017035)

Remote sensing extraction of fruit tree planting area based on Sentinel-2 multi-temporal images

ZHOU Xinxing1(), ZHAO Lin1,*(), ZHANG Wenjie1, TAN Changwei2, LI Gangbo1, SHI Mengyun1, ZHANG Ting1, YANG Feng1   

  1. 1. Xuzhou Institute of Agricultural Sciences in Xuhuai Area of Jiangsu, Xuzhou 221121, Jiangsu, China
    2. Agricultural College of Yangzhou University, Yangzhou 225009, Jiangsu, China
  • Received:2021-10-18 Online:2022-12-25 Published:2022-12-26
  • Contact: ZHAO Lin

摘要:

为提取果树的空间分布信息,以果树生长期内不同月份的Sentinel-2多光谱遥感影像为数据源,以大沙河流域果树为研究对象,通过分析不同月份的光谱信息得出最佳监测时期,并在此基础上,选择不同时期的5种植被指数[归一化植被指数(NDVI)、比值植被指数(RVI)、增强型植被指数(EVI)、结构密集型色素指数(SIPI)和归一化水指数(NDWI)],结合机器学习技术构建决策树提取模型。结果发现,3、4、7、8月份的影像适于果树面积提取。通过Feature_importances_属性筛选出贡献度高的不同时期的植被指数作为输入特征,结合超参数学习曲线和网格搜索技术确定决策树模型的Max_depth和Min_samples_leaf参数分别为5和10时模型的效果最佳。参数调整后绘制决策树模型,模型在训练集和测试集上的精度分别达到了0.919 4和0.875 1。提取结果表明,研究区内的果树主要种植在大沙河两岸,东部与西北部的果树种植地块较为零碎,总的果树种植面积为6 838 hm2。在验证样本的基础上,通过混淆矩阵计算提取结果的精度,结果显示,Kappa系数为0.87,果树种植区提取的用户精度和制图精度分别为92.91%和90.77%。结果说明,本文所提出的方法适用于大区域果树的遥感提取,可为基于中高分辨率遥感影像的果树种植区监测提供有效的技术手段。

关键词: 果树种植区, 遥感, 决策树, 机器学习

Abstract:

In order to extract the spatial distribution information of fruit trees, the Dasha River Basin was selected as the study area, and the Sentinel-2 multispectral remote sensing images of different months were used as the data source. The best monitoring period was obtained by analyzing the spectral information of different months. On this basis, five vegetation indices of normalized differential vegetation index (NDVI), ratio vegetation index (RVI), enhance vegetation index (EVI), structure intensive pigment index (SIPI) and normalized difference water index (NDWI) in different periods were selected to construct a decision tree extraction model combined with machine learning technology. The results showed that the images in March, April, July, and August were suitable for the extraction of fruit tree planting area. The vegetation indices in different months with high contribution were selected through the attribute of Feature_importances_, as the input feature. Based on the combination of the hyperparameter learning curve and grid search technology, parameters of Max_depth and Min_samples_leaf were determined as 5 and 10, respectively, as the model effect was the best under these parameters. After the adjustment of parameters, the decision tree model was drawn, and the accuracies of the model on the training set and the test set were 0.919 4 and 0.875 1, respectively. The extraction results showed that fruit tree planting area was mainly in the banks of the Dasha River, and the planting plots of fruit trees in the east and northwest were relatively fragmented. The total planting area was 6 838 hm2. On the basis of the verification sample, the accuracy of the extraction results was calculated by the confusion matrix. The Kappa coefficient was 0.87, and the user accuracy and mapping accuracy of fruit tree planting area extraction were 92.91% and 90.77%, respectively. Thus, the proposed method could be applied to remote sensing extraction of fruit trees in a large area, and could provide effective technical means for the monitor of fruit tree planting areas with medium and high resolution remote sensing images.

Key words: fruit tree planting area, remote sensing, decision tree, machin learning

中图分类号: