基于多物候特征指数相关性迁移的冬小麦多年份分布信息识别

Identification of Multi-year Distribution Information of Winter Wheat Based on Correlation Transfer of Multi-phenological Feature Indices

摘要:

遥感作物识别中，样本数据对识别精度有重要影响，而大区域多年份获取样本数据是一项十分繁琐的工作。为减少逐年样本采集工作量，提高作物识别效率，提出一种基于多物候特征指数的样本迁移策略。使用2019年焦作市冬小麦分布图，利用多物候特征指数时间序列曲线相关性迁移生成2020、2021年高质量样本数据，并利用随机森林机器学习方法实现了2020、2021年焦作市冬小麦自动高效识别。结果表明：利用提出的样本迁移策略获取样本数据，当显著性水平达到0.001时，2年冬小麦识别总体精度均在94%以上，Kappa系数均在0.91以上，各县（市）识别面积与统计面积决定系数（R2）达到0.957，均方根误差（RMSE）为20.16km2。与单一植被指数时间序列曲线相关性迁移方法相比，该方法使2020年与2021年识别总体精度分别提高1.32、2.27个百分点，Kappa系数分别提升0.022、0.037，2年各县（市）识别面积与统计面积R2提高0.026，RMSE减少20.1%。此外，将该迁移策略应用于新乡市与鹤壁市，冬小麦识别总体精度均在92%以上，识别面积与统计面积的R2也达到0.92。表明提出的样本迁移策略在跨时间与跨地域中均表现较好，可为进一步快速、精准获取大区域长时序作物分布信息提供思路与技术支撑。

关键词:冬小麦;分布信息识别;Sentinel-2;时间序列;迁移学习;多物候特征指数

Abstract:

In remote sensing crop identification, the quality of sample data significantly influences the accuracy of identification. However, collecting sample data for multiple years in large regions is a laborious task. To reduce the workload of annual sample collection and improve crop identification efficiency, a sample transfer strategy was proposed based on multi-phenological feature indices. Utilizing the winter wheat distribution map of Jiaozuo City in 2019, high-quality sample data for 2020 and 2021 were generated by the correlation of multi-phenological feature index time series curves. Then the random forest machine learning method was employed to achieve automatic and efficient identification of winter wheat in Jiaozuo City for 2020 and 2021. The results indicated that by employing the proposed sample migration strategy to obtain the sample data, the overall accuracy of winter wheat identification in both years exceeded 94% when the correlation reached 0.001 at the significance level. Additionally, the Kappa coefficient was above 0.91. The coefficient of determination (R2) between the identified area and the statistical area for each county (city) reached 0.957, with root mean square error (RMSE) of 20.16km2. This demonstrated the effectiveness and precision of the proposed method in winter wheat identification. Compared with the method of transferring correlation based on single vegetation index time series curves, the overall accuracy of the method in 2020 and 2021 was improved by 1.32 percentage points and 2.27 percentage points, respectively. The Kappa coefficients were increased by 0.022 and 0.037, respectively, and the R2 between the identified areas and the statistical areas of each county was increased by 0.026 over the two years, the RMSE was decreased by 20.1%. Furthermore, when applying this transfer strategy to Xinxiang City and Hebi City, the overall accuracy of winter wheat identification exceeded 92%, and the R2 between the identified areas and the statistical areas was 0.92. The research result demonstrated that the proposed sample transfer strategy performed well across both time and space, providing insights and technical support for rapidly and accurately obtaining longterm crop distribution information in large regions.

Key words:winter wheat;distribution information identification;Sentinel-2;time series;transfer learning;multi-phenological feature indices