收藏设为首页 广告服务联系我们在线留言

干旱区地理 ›› 2024, Vol. 47 ›› Issue (6): 1004-1014.doi: 10.12118/j.issn.1000-6060.2023.382

• 生物与环境 • 上一篇    下一篇

基于RF分类调优和SNIC聚类的新疆红枣种植区遥感提取

赵国兵1,2(), 郑江华1,3, 王蕾2(), 高健2, 罗磊2, 尼格拉·吐尔逊1, 韩万强1, 关靖云1   

  1. 1.新疆大学地理与遥感科学学院,新疆 乌鲁木齐 830046
    2.新疆林业科学院现代林业研究所,新疆 乌鲁木齐 830063
    3.新疆绿洲生态重点实验室,新疆 乌鲁木齐 830046
  • 收稿日期:2023-07-23 修回日期:2023-10-07 出版日期:2024-06-25 发布日期:2024-07-09
  • 通讯作者: 王蕾(1981-),女,研究员,主要从事林业遥感研究. E-mail: 18699135569@163.com
  • 作者简介:赵国兵(1999-),男,硕士研究生,主要从事林业遥感研究. E-mail: zhaoguobing9910@163.com
  • 基金资助:
    基于“空天地”多源遥感监测技术的林果资源数据体系建设(20222101536);2023年中央财政林草科技推广示范项目—新疆和田地区林果资源监测技术典型示范与推广(新[2024]TG15)

Remote sensing extraction of jujube planting area in Xinjiang based on RF classification optimization and SNIC clustering

ZHAO Guobing1,2(), ZHENG Jianghua1,3, WANG Lei2(), GAO Jian2, LUO Lei2, Nigela TUERXUN1, HAN Wanqiang1, GUAN Jingyun1   

  1. 1. College of Geography and Remote Sensing Sciences, Xinjiang University, Urumqi 830046, Xinjiang, China
    2. Institute of Modern Forestry, Xinjiang Academy of Forestry Sciences, Urumqi 830063, Xinjiang, China
    3. Xinjiang Key Laboratory of Oasis Ecology, Urumqi 830046, Xinjiang, China
  • Received:2023-07-23 Revised:2023-10-07 Published:2024-06-25 Online:2024-07-09

摘要:

论文旨在快速提取新疆红枣种植区分布信息和种植面积,为预测产量、价格,巩固脱贫攻坚成果和助力乡村振兴提供数据支持。基于Google Earth Engine云平台,快速获取覆盖全疆的Sentinel-1雷达影像、Sentinel-2光学影像及SRTM地形数据,从中提取光谱、纹理、地形等44个特征并进行特征优选过程,在对随机森林分类器进行超参数调优后,得到新疆2021年10 m分辨率红枣种植区空间分布图,运用超像素聚类的方法对全疆主要红枣种植区域进行分类后处理及分区统计,最终得到全疆红枣种植面积。结果表明:(1) 通过基于简单非迭代聚类算法进行分类处理,得到全疆红枣种植面积为4253 km2,其主要分布在南疆的阿克苏、喀什、和田地区、巴音郭楞蒙古自治州和东疆的吐鲁番市、哈密市等地。(2) 对随机森林分类器进行超参数调优后,能够有效提高提取精度,基于混淆矩阵计算的平均总体分类精度为0.86,平均Kappa系数为0.82,红枣提取的生产者精度为0.87,用户精度为0.80。(3) Sentinel-1极化波段特征在红枣信息提取中占据重要地位,光谱特征和纹理特征次之。结合多源遥感数据能够快速提取新疆红枣种植区分布与面积信息,对推动该地区农业现代化、资源保护和经济发展具有重要意义。

关键词: Google Earth Engine, Sentinel-1/2, 红枣, 特征优选, 随机森林, 超像素聚类

Abstract:

This study aims to efficiently extract the distribution information and planting area of jujube crops in Xinjiang, China, providing essential data support for predicting yield and price, consolidating poverty alleviation achievements, and aiding rural revitalization. Utilizing the Google Earth Engine cloud platform, this research accesses Sentinel-1 radar images, Sentinel-2 optical images, and SRTM terrain data covering Xinjiang. From these data, 44 features including spectral, textural, and terrain attributes are extracted, followed by a feature selection process. The optimized random forest classifier, after hyperparameter tuning, produces a spatial distribution map of Xinjiang’s jujube planting areas with a 10 m resolution for the year 2021. Superpixel clustering method further processes the major jujube planting areas to determine the exact planting extents. The findings are as follows: (1) Employing a simple non-iterative clustering algorithm for classification and post-processing, the identified jujube cultivated area in Xinjiang spans 4253 km², predominantly located in the southern regions of Aksu, Kashgar, Hotan Prefectures, and Bayingolin Mongol Autonomous Prefecture, as well as Turpan and Hami Cities in the east. (2) The accuracy of feature extraction is significantly enhanced through hyperparameter optimization of the random forest classifier, yielding an average overall classification accuracy of 0.86, a Kappa coefficient of 0.82, a producer accuracy for jujube extraction of 0.87, and a user accuracy of 0.80, as assessed via the confusion matrix. (3) Features from the Sentinel-1 polarization band are crucial for jujube information extraction, supplemented effectively by spectral and textural features. Leveraging multisource remote sensing data, this method facilitates rapid acquisition of distribution and area data for jujube planting in Xinjiang, markedly benefiting agricultural modernization, resource conservation, and regional economic development.

Key words: Google Earth Engine, Sentinel-1/2, jujube, feature optimization, random forest, super-pixel clustering