【目的】基于无人机高光谱影像构建油茶叶片氮磷钾(NPK)含量的估测模型并评价模型精度,快速、实时、准确地测算油茶叶片的NPK含量以精准开展施肥管理,提高油茶果的产量。【方法】以广东省河源市的美林湖油茶样地为研究对象,基于成熟油茶叶片的NPK含量和无人机高光谱影像(大疆DJIM600无人机搭载的GaiaSKY-mini2传感器),构建基于偏最小二乘法(PLSR)和随机森林(RF)的油茶叶片NPK含量估算模型,并采用留一交叉验证法,以均方根误差(RMSE)、调整均方根误差(adjRMSE)和决定系数(R²)评价模型精度。【结果】在相同的RMSE水平下,基于RF构建的叶片氮磷(NP)含量估算模型表现更优,如估算N、P含量的RMSE分别为0.70、0.09 mg/g时,RF模型的R²分别为0.84、0.79,PLSR模型R²分别为0.74、0.85;而油茶叶片K含量估算效果则以PLSR模型更优,其RMSE和R²分别为0.30 mg/g和0.86,RF模型的RMSE和R²分别为0.35 mg/g和0.81。估算叶片N、P含量的RF模型中,4个权重最高的高光谱影像反射率分别位于637.7、497.6、661.5、468.3 nm波段(N含量)和734.0、397.7、583.7、587.0 nm波段(P含量);而估算叶片K含量的PLSR模型中,4个权重最高的高光谱影像反射率位于394.5、397.7、423.2、400.9 nm波段。【结论】基于机器学习模型(PLSR模型和RF模型),在使用无人机高光谱影像反演油茶叶片NPK含量和筛选特征高光谱波段方面比相关性分析方法更有优势。本研究提供了一种高精度、高频率、快速无损估算油茶叶片NPK含量的方法,可为油茶施肥管理和提高产量提供有效的技术支持。

【Objective】This study aims to develop estimation models for leaf nitrogen (N), phosphorus (P), and potassium (K) contents in Camellia oleifera based on unmanned aerial vehicle (UAV) hyperspectral imagery and to evaluate their accuracy. The goal was to achieve rapid, real-time, and accurate monitoring of leaf NPK contents to support precise fertilization management and enhance the yield of oil tea crops.【Method】Using a mature C. oleifera plantation in Meilinhu, Heyuan City, Guangdong Province, as the study site, leaf NPK measurements were integrated with UAV hyperspectral data acquired by a GaiaSKY-mini2 sensor mounted on a DJI Matrice 600 (M600) platform. Estimation models for leaf NPK content were constructed using partial least squares regression (PLSR) and random forest (RF) algorithms. Model performance was evaluated using leave-one-out cross-validation, with the coefficient of determination (R²), root mean square error (RMSE), and adjusted RMSE (adjRMSE) serving as the primary accuracy metrics.【Result】At comparable RMSE levels, the RF algorithm demonstrated superior performance in estimating leaf nitrogen and phosphorus (N,P) content. Specifically, when the RMSE for N and P estimation was 0.70 mg/g and 0.09 mg/g, respectively, the corresponding R² values for the RF models were 0.84 and 0.79, while those for the PLSR models were 0.74 and 0.85. Conversely, the PLSR model outperformed RF in estimating leaf potassium (K) content, yielding an RMSE of 0.30 mg/g and an R² of 0.86, compared to an RMSE of 0.35 mg/g and an R² of 0.81 for RF. Feature band analysis revealed that the most influential hyperspectral bands for estimating N content via RF were centered at 637.7, 497.6, 661.5, and 468.3 nm, while those for P content were at 734.0, 397.7, 583.7, and 587.0 nm. For K content estimation using PLSR, the key bands were identified at 394.5, 397.7, 423.2, and 400.9 nm.【Conclusion】Machine learning models (PLSR and RF) demonstrate significant advantages over traditional correlation analysis in the inversion of leaf NPK content and the selection of characteristic spectral bands using UAV hyperspectral imagery. This study provides a high-precision, high-frequency, rapid, and non-destructive method for monitoring the nutritional status of Camellia oleifera, offering effective technical support for optimized fertilization management and yield improvement.