基于可解释机器学习的冬油菜叶片氮、磷、钾含量遥感估测

Linking Leaf Phenotypes to Spectral Characteristics for Remote Estimation of Leaf Nitrogen, Phosphorous, and Potassium Concentration in Winter Oilseed Rape (Brassica napus L.)

摘要:

通过高光谱遥感估测作物叶片养分含量，关键在于识别对养分缺乏敏感的波长范围。以往研究主要关注光谱数据与养分含量间的相关性，忽略了敏感波段与养分胁迫下表型变化间的关系，从而导致波段选择的解释性不足。本文提出了一种基于辐射传输模型和后解释算法的光谱特征选择方法，采用极端梯度提升模型结合后解释算法（Shapley additive explanations, SHAP）进行敏感表型筛选，使结果具有可解释性。随后将叶片光谱与表型变化相结合，以筛选对冬油菜氮、磷和钾胁迫敏感的光谱特征，并利用筛选的光谱特征估测油菜叶片氮、磷、钾含量，从而验证波段筛选有效性。研究结果表明，蛋白质SHAP值最高，表明其是氮养分缺乏的关键表型，花青素和类胡萝卜素次之，因此选择对这3个表型敏感的400~720 nm和1 140~2 250 nm间43个波段用于估算叶片氮含量（Leaf nitrogen concentration, LNC）。对于磷缺乏，SHAP值最高的花青素是最重要表型，其次是蛋白质和叶绿素，对应选择了400~810 nm以及1 140~2 250 nm间50个波段用于估算叶片磷含量（Leaf phosphorus concentration, LPC）。此外，类胡萝卜素是对钾缺乏最敏感的表型，叶绿素和等效水厚度的SHAP值也较高，故选择与其密切相关的450~810 nm和1 140~2 480 nm间32个波段估算叶片钾含量（Leaf potassium concentration, LKC）。根据筛选的波段构建随机森林模型估测LNC、LPC和LKC，模型在验证集上实现了对LNC、LPC和LKC的准确估算，其中LNC估算精度最高（决定系数（R2）为0.86，均方根误差（RMSE）为0.37%，归一化均方根误差（NRMSE）为0.11），其次为LPC（R2=0.83, RMSE为0.04%, NRMSE为0.10），而LKC精度相对较低（R2=0.79, RMSE为0.35%, NRMSE为0.12）。研究结果突出体现了敏感光谱波段选择在叶片养分含量估算中的重要性，并揭示了这些波段与氮、磷、钾胁迫下表型特征间的关系。

Abstract:

The estimation of crop leaf nutrient content through hyperspectral remote sensing necessitates the identification of wavelength bands sensitive to nutrient deficiencies. Previous studies focused on correlating spectral data with nutrient content, neglecting the relationship between sensitive bands and phenotypic changes under nutrient stress, leading to less explainable band selection. Here, a spectral feature selection method was proposed, integrating a radiative transfer model with a post?explanation algorithm, specifically using the extreme gradient boosting model combined with Shapley additive explanations (SHAP) to enhance the interpretability of sensitive phenotypic selection. By linking leaf spectra with phenotypic changes, spectral features sensitive to nitrogen (N), phosphorus (P), and potassium (K) deficiencies in winter oilseed rape (Brassica napus L.) were identified. These selected spectral features were subsequently used to estimate leaf nitrogen concentration (LNC), leaf phosphorus concentration (LPC), and leaf potassium concentration (LKC), thereby validating the effectiveness of the band selection process. The results revealed that protein, with the highest SHAP value, was the key phenotypic indicator for N deficiency, followed by anthocyanin and carotenoid, therefore, prompting the selection of 43 bands between 400~720 nm and 1 140~2 250 nm for estimating LNC. For P deficiency, anthocyanin, with the highest SHAP value, was the most critical phenotype, followed by protein and chlorophyll, resulting in the selection of 50 bands between 400~810 nm and 1 140~2 250 nm for estimating LPC. Furthermore, carotenoid was identified as the most sensitive phenotype to K deficiency, with chlorophyll and equivalent water thickness also showing high SHAP values, leading to the selection of 32 bands between 450~810 nm and 1 140~2 480 nm for estimating LKC. The selected bands were used to construct random forest models to estimate LNC, LPC, and LKC. These models demonstrated high accuracy on the validation dataset. The highest performance for LNC, with a coefficient of determination (R2) of 0.86, a root mean square error (RMSE) of 0.37%, and a normalized root mean square error (NRMSE) of 0.11, followed by LPC (R2 of 0.83, RMSE of 0.04%, and NRMSE of 0.10), and LKC (R2 of 0.79, RMSE of 0.35%, and NRMSE of 0.12). Results underscored the significance of selecting sensitive wavelength bands for estimating leaf nutrient content and illuminated their relationship with phenotypes under N, P, and K deficiencies.