基于可解释性机器学习的水稻氮素浓度无人机高光谱反演方法研究

Hyperspectral inversion methods for rice nitrogen concentration based on explainable machine learning using UAV

关键词:  无人机  高光谱  SHAP分析  氮素  水稻

Key words:  unmanned aerial vehicle（UAV）  hyperspectral  SHAP analysis  nitrogen  rice

摘要:

【目的】 水稻氮素营养状况直接决定稻谷的品质与产量，准确获取氮素浓度是规模化水稻生产中实现氮肥精准施用的关键依据。当前基于机器学习方法构建无人机高光谱数据与氮素浓度间的统计模型，是监测水稻氮素营养状况的技术路径之一。【方法】 为解决水稻氮素浓度无人机高光谱反演模型可解释性不足的问题，文章通过光谱预处理、特征筛选与多模型对比构建水稻氮素浓度反演模型。首先利用Savitzky-Golay方法对原始光谱数据进行平滑，并以3 nm为间隔进行重采样，采取连续投影算法（SPA）提取水稻氮素特征波段，以极限学习机（ELM）、随机森林（RF）和反向传播神经网络（BPNN）为建模方法，构建无人机高光谱氮素反演模型，并基于SHAP（SHapley Additive exPlanations）可解释性方法对最优反演模型每个指标的贡献度进行揭示。【结果】 （1）SPA筛选出特征波段15个，分别为472 nm、547 nm、613 nm、787 nm、874 nm、976 nm、418 nm、481 nm、592 nm、862 nm、895 nm、907 nm、934 nm、979 nm和994 nm；（2）ELM模型反演效果最优，测试集R²为0.771、RMSE为0.478%、NRMSE为0.131%，显著优于RF和BPNN模型；（3）SHAP揭示了最优反演模型下592 nm、862 nm和481 nm等波段对氮素浓度反演结果的贡献值分别为2.097、1.764和1.431。【结论】 基于ELM构建的水稻氮素浓度无人机高光谱反演模型能够较为高效且准确地获得水稻氮素浓度信息，为水稻的精准施肥决策提供数据支撑。

Abstract:

[Purpose] The nitrogen nutritional status of rice directly determines grain quality and yield. Accurate determination of nitrogen concentration serves as a critical basis for achieving precise nitrogen fertiliser application in large-scale rice production. Currently，constructing statistical models linking unmanned aerial vehicle（UAV）hyperspectral data to nitrogen concentration using machine learning methods represents one technical pathway for monitoring rice nitrogen nutrition.[Method] To address the lack of interpretability in UAV hyperspectral inversion models for rice nitrogen concentration，this study constructed an inversion model through spectral preprocessing，feature selection，and multi-model comparison. First，raw spectral data was smoothed using the Savitzky-Golay method and resampled at 3 nm intervals. The successive projections algorithm（SPA）was employed to extract characteristic nitrogen bands，whilst the extreme learning machine（ELM），random forest（RF），and backpropagation neural network（BPNN）were used as modelling approaches to construct UAV hyperspectral nitrogen inversion models. The shapley additive explanations（SHAP）method was employed to reveal the contribution of each indicator to the optimal inversion model.[Result] （1）The SPA identified 15 characteristic spectral bands：472 nm，547 nm，613 nm，787 nm，874 nm，976 nm，418 nm，481 nm，592 nm，862 nm，895 nm，907 nm，934 nm，979 nm，and 994 nm；（2）The ELM model demonstrated optimal inversion performance，achieving an R² of 0.771，an RMSE of 0.478%，and an NRMSE of 0.131% on the test set，significantly outperforming both RF and BPNN models；（3）The SHAP analysis revealed that the contribution values of the 592 nm，862 nm，and 481 nm bands were 2.097，1.764，and 1.431 respectively in the optimal inversion model's nitrogen concentration results.[Conclusion] The UAV hyperspectral inversion model for rice nitrogen concentration based on ELM can efficiently and accurately obtain nitrogen concentration information，providing data support for precision fertilization decisions in rice cultivation.