欢迎访问《中国农学通报》，
中国农学通报

中国农学通报 ›› 2024, Vol. 40 ›› Issue (25): 155-164.doi: 10.11924/j.issn.1000-6850.casb2023-0760

• 农业信息·科技教育 • 上一篇    

基于多源卫星的黄淮海平原冬小麦种植丰度定量评估

王锦杰1(), 陈昊2(), 庞礴1, 周航1, 沈伟1, 颜雅琼1, 徐敏3   

  1. 1 宿迁市气象局，江苏宿迁 223800
    2 江苏省气象台，南京 210019
    3 江苏省气候中心，南京 210019
  • 收稿日期:2023-11-08 修回日期:2024-06-18 出版日期:2024-09-05 发布日期:2024-08-27
  • 通讯作者:
    陈昊，男，1986年出生，高级工程师，博士，主要从事微波毫米波辐射传输理论与环境气象预报。通信地址：210008 江苏省南京市建邺区雨顺路6号 江苏省气象台，E-mail：。
  • 作者简介:

    王锦杰，男，1992年出生，工程师，硕士，主要从事农业遥感研究。通信地址：223800 江苏省宿迁市宿城区洪泽湖路722号，E-mail：。

  • 基金资助:
    江苏省气象局揭榜挂帅项目“保障粮食安全的气象灾害风险预警评估技术研究”(KZ202302); 江苏省气象局青年基金项目“基于机器学习的大范围冬小麦种植丰度定量评估方法”(KQ202420); 江苏省第六期“333人才”培养支持项目“小麦赤霉病流行新特征和全程防控精细化气象预测技术”; 江苏省气象局青年基金项目“淮北地区气象干旱到农业生态干旱的动态传播特征及其驱动力研究”(KQ202330)

Quantitative Assessment of Winter Wheat Planting Abundance in Huang-Huai-Hai Plain Based on Multi-source Satellite Data

WANG Jinjie1(), CHEN Hao2(), PANG Bo1, ZHOU Hang1, SHEN Wei1, YAN Yaqiong1, XU Min3   

  1. 1 Suqian Meteorological Bureau, Suqian, Jiangsu 223800
    2 Jiangsu Meteorological Observatory, Nanjing 210019
    3 Jiangsu Provincial Climate Center, Nanjing 210019
  • Received:2023-11-08 Revised:2024-06-18 Published:2024-09-05 Online:2024-08-27

RichHTML

4

PDF (PC)

9

摘要：

针对大范围冬小麦种植丰度定量评估和种植面积测量业务化中存在的，高分辨率影像覆盖能力较低难以在大空间范围内推广应用，与中分辨率影像提取精度较低之间相互制约的现实问题，选择均匀分布于黄淮海平原的6个Sentinel-2条带位置为试验区，通过分别构建随机森林分类模型提取Sentinel-2的冬小麦种植区域，并将Sentinel-2冬小麦种植区域合成为250 m空间分辨率的种植丰度，结合时序MODIS NDVI训练随机森林回归模型，预测得到黄淮海平原冬小麦种植丰度，从而实现大范围冬小麦种植丰度定量评估和种植面积测量。相比传统MODIS NDVI时序数据提取冬小麦种植区域，还需额外进行混合像元分解后才能得到种植丰度，本研究使用随机森林回归方法直接获得了每个像元的种植丰度，省去了混合像元分解步骤。训练的各条带位置随机森林分类模型，F1 score达0.9983以上，当训练集样本量占总样本量的2%以上时随机森林回归模型趋于稳定，当样本量占比达50%时模型最适宜使用，R2达0.8140，样本量占比达90%时，回归模型R2达到最大值为0.8162。使用模型测量冬小麦种植丰度和种植面积分别能够达到Sentinel-2精度的91%和99%以上，满足了大范围冬小麦种植丰度定量评估和种植面积测量的业务化精度要求。

关键词: 冬小麦, 中高分辨率结合, 种植丰度, 随机森林, 分类, 回归

Abstract:

To overcome the challenge posed by the mutual restriction between the limited coverage capacity of high-resolution images, which hinders their promotion and application on large spatial scales, and the relatively low extraction accuracy of medium-resolution images in quantitatively assessing winter wheat planting abundance and measuring planting areas over extensive regions, this study conducted experiments in the Huang-Huai-Hai Plain. Six evenly distributed Sentinel-2 tiles covering a vast spatial area served as the experimental region. The winter wheat planting area, derived from Sentinel-2 data, was converted into planting abundance with a spatial resolution of 250 meters. Subsequently, planting abundance data and time-series MODIS NDVI were combined to train a random forest regression model, aiming to achieve quantitative assessments of winter wheat planting abundance and measurements of planting areas on a large scale. In contrast to traditional methods relying on MODIS NDVI time-series data for extracting winter wheat planting areas, which necessitate additional mixed image decomposition to ascertain planting abundance, this study utilized the random forest regression approach to directly ascertain the planting abundance of each pixel, thereby eliminating the mixed image decomposition step. The trained random forest classification model exhibited an F1 score exceeding 0.9983 across different tiles. The random forest regression model tended to stabilize when the training set sample size comprised over 2% of the total sample size. The model proved most suitable for use at a 50% sample size, yielding an R2 value of 0.8140. At a 90% sample size, the regression model achieved a maximum R2 of 0.8162. The random forest regression models predicted winter wheat planting abundance and area with accuracies exceeding 91% and 99% of Sentinel-2’s accuracy, respectively. This meets the operational accuracy requirements for large-scale quantitative evaluations of winter wheat planting abundance and area measurements.

Key words: winter wheat, combination of medium and high resolution, planting abundance, random forest, classification, regression