融合时序RGB-多光谱的油菜冠层叶片氮监测模型与追肥处方图构建方法

Integrated Time-series RGB-Multispectral Approach for Canopy Leaf Nitrogen Monitoring Model and Topdressing Prescription Mapping in Rapeseed

摘要:

氮元素含量是表征作物生长状况的重要指标,氮含量监测对作物生长发育和施肥管理决策具有重要意义。针对传统氮含量监测精度不足、成本高、空间覆盖有限等问题,提出了一种融合时序RGB-多光谱的油菜氮监测与追肥处方图构建方法。首先利用大疆Mavic 3M 型多光谱无人机,连续两年采集试验田油菜冠层叶片氮含量值和对应油菜生育关键期RGB 与多光谱时序遥感影像,建立了冠层RGB与多光谱特征组合;采用相关性分析方法筛选对油菜叶氮敏感的特征,通过最大相关和最小冗余(Maximum relevance and minimum redundancy, mRMR)算法对筛选后特征值进行重要性排序,将排序后特征值依次利用随机森林(Random forest, RF)、岭回归(Ridge regression,RR)、支持向量回归(Support vector regression, SVR)、梯度提升回归(Gradient boosting regression, GBR)模型获取最优时序特征组合,用于构建苗期(六叶期)、越冬期和蕾薹期氮监测模型;根据种植区域的油菜氮含量分布图,对田块进行栅格划分并决策追肥量,生成了油菜追肥处方图。数据分析结果显示:基于RF 模型、融合多生育期时序信息的RGB与多光谱最优特征组合(LCI、MSAVI、NDVI、ExG、TH)监测效果最佳,可同时监测六叶期、越冬期和蕾薹期氮含量,R2分别为0.825、0.800和0.806,RMSE分别为0.210%、0.231%和0.217%;相较于仅使用单一生育期数据建模,融合时序信息后六叶期、越冬期和蕾薹期的R2分别提高了0.100、0.036和0.061。通过分析无人机20~50m不同飞行高度对监测精度的影响,确定了40m为油菜氮含量监测的适宜飞行高度,六叶期、越冬期和蕾薹期的R2分别为0.792、0.770和0.787。研究表明,融合时序RGB多光谱可有效监测油菜冠层叶片氮含量,研究结果可为油菜作物氮营养精准管理与追肥处方决策提供理论依据和技术参考。

Abstract:

Nitrogen content is a crucial indicator for assessing crop growth status, and its monitoring plays an important role in crop growth, development, and fertilization management decisions. Aiming to address the limitations of traditional nitrogen monitoring methods, including insufficient accuracy, high costs, and limited spatial coverage, a method was proposed for monitoring nitrogen content in rapeseed and constructing topdressing prescription maps by integrating time-series RGB and multispectral features. Firstly, a DJI Mavic 3M multispectral drone was used to collect canopy leaf nitrogen content values and corresponding time-series RGB and multispectral remote sensing images of rapeseed during key growth stages over two consecutive years, and a combination of canopy RGB and multispectral features was established. Correlation analysis was employed to identify the features sensitive to rapeseed leaf nitrogen content, and the maximum relevance and minimum redundancy (mRMR) algorithm was applied to rank the importance of the selected features. The ranked features were then sequentially input into random forest (RF), ridge regression (RR), support vector regression (SVR), and gradient boosting regression (GBR) models to obtain the optimal time-series feature combination for constructing nitrogen content monitoring models at the seedling stage (six-leaf stage), overwintering stage, and bolting stage. Based on the nitrogen content distribution map of the rapeseed planting area, the field was divided into grids to determine topdressing rates, and a rapeseed topdressing prescription map was generated. The data analysis results showed that the optimal feature combination based on the RF model and integrating multi-growth-stage time-series information (LCI, MSAVI, NDVI, ExG, TH) achieved the best monitoring performance, with R2 values of 0.825, 0.800, and 0.806 and RMSE values of 0.210%, 0.231%, and 0.217% for the six-leaf stage, overwintering stage, and bolting stage, respectively. Compared with modeling using only single-growth-stage data, the integration of time-series information improved the R2 values for the six-leaf stage, overwintering stage, and bolting stage by 0.100, 0.036, and 0.061, respectively. Additionally, by analyzing the effect of different UAV flight altitudes (20~50 m) on monitoring accuracy, 40m was determined as the optimal flight altitude for rapeseed nitrogen content monitoring, with R2 values of 0.792, 0.770, and 0.787 at the six-leaf stage, overwintering stage, and bolting stage, respectively. The research result demonstrated that integrating time-series RGB and multispectral features could effectively monitor nitrogen content in rapeseed canopy leaves. The findings can provide a theoretical basis and technical reference for precision nitrogen management and topdressing prescription decisions in rapeseed crops.