基于高斯回归分析的水稻氮素敏感波段筛选及含量估算

水稻氮素含量的准确监测是稻田精准施肥的重要环节，水稻叶片氮素含量发生变化会引起叶片、冠层的光谱发射率发生变化，高光谱遥感是目前作物氮素无损监测的关键技术之一。以2018年-2019年湖北监利两年水稻氮肥试验为基础，分别获取水稻分蘖期、拔节期、孕穗期、扬花期、灌浆期五个生育期水稻叶片和冠层两个尺度的高光谱反射率数据及对应的叶片氮素含量数据，利用单波段原始光谱和一阶导数光谱的相关性分析、高斯过程回归（GPR）等方法筛选水稻全生育期叶片及冠层尺度氮素敏感波段。针对敏感波段，利用单波段回归分析、随机森林（RF）、支持向量回归（SVR）、高斯过程回归-随机森林（GPR-RF）、高斯过程回归-支持向量回归（GPR-SVR）和GPR构建水稻氮素监测模型，并进行精度对比，以确定水稻叶片在各生育期的氮素估算最佳模型。结果表明：GPR筛选的敏感波段符合水稻氮素含量及光谱变化的规律。相同条件下，叶片模型精度整体高于冠层模型。相关性分析模型中，叶片尺度原始光谱模型更好，冠层尺度刚好相反，冠层一阶导数光谱可以减弱稻田背景噪声的影响。其中，叶片最佳模型建模集R2为0.79，验证集R2为0.84；冠层最佳模型建模集R2为0.80，验证集R2为0.77。与相关性回归分析模型相比，机器学习模型受生育期影响小（R2>0.80，NRMSE<10%）。其中，RF比SVR更适合对GPR敏感波段建模，GPR-RF模型可以用1.5%左右的波段达到RF模型使用全部波段的精度。五种方法中，GPR模型对生育期敏感度最低、叶片及冠层尺度效果都很好(R2>0.94，NRMSE<6%)。且与其他四种机器学习方法相比，GPR模型可有效提高冠层氮素含量估算的精度和稳定性（R2增加0.02，NRMSE降低1.2%）。GPR方法可为筛选作物氮素高光谱敏感波段、反演各生育期叶片及冠层氮素含量提供方法参考。

Sensitive Bands Selection and Nitrogen Content Monitoring of Rice Based on Gaussian Regression Analysis

Accurate detection of rice nitrogen content is an important aspect of precision fertilization in rice fields. Nitrogen content variation of rice leaves will cause changes in emissivity of leaves and canopy. Hyperspectral remote sensing is one of the key technologies for non-destructive monitoring of crop nitrogen. This study focuses on the study of nitrogen content monitoring and the sensitive band’s selection through machine learning methods based on 2-year rice nitrogen fertilization experiments carried out in Jianli Hubei during 2018-2019. Hyperspectral reflectance spectral data at the leaf and canopy level and the corresponding leaf nitrogen content data were collected at rice tillering, jointing, booting, flowering and filling stage, respectively. Correlation analysis and Gaussian process regression (GPR) were used to select nitrogen sensitive bands for raw spectra and first-order derivative reflectance (FDR) spectra in rice leaves and canopy level. The nitrogen content estimation models were then constructed through single-band regression analysis, Random Forest (RF) and Support Vector Regression (SVR) method for rice raw spectra data. The Gaussian Process Regression-Random Forest (GPR-RF), Gaussian Process Regression- Support Vector Regression (GPR-SVR), and GPR method were also used to construct the nitrogen estimation model for the nitrogen-sensitive selection bands. The results showed that the GPR method’s sensitive bands were consistent with variation rule of the nitrogen content and spectral changes in rice. The leaf-level model’s over all accuracy was higher than that of the canopy-level model under the same conditions while using FDR spectra was more accurate at canopy level for it could attenuate the effect of background noise in the rice field. R2 of calibration datasets and validation sets are 0.79 and 0.84 at the leaf level, while 0.80 and 0.77 at canopy level. Compared with the correlation regression model, the machine learning methods were less affected by rice growth stages (R2>0.80, NRMSE<10%). RF was more suitable than SVR for modeling GPR-selection nitrogen sensitive bands, and the GPR-RF model can use about 1.5% of the bands to reach the accuracy of the RF model using all the bands. The GPR model works well on nitrogen estimation through nitrogen -sensitive bands at leaf and canopy level, not only for the full-growth stage but also for the single-growth stage(R2>0.94, NRMSE<6%). Besides, compared with the other four machine learning methods, the GPR model can improve the accuracy and stability of the estimation of nitrogen content at the canopy level that R2 increased by 0.02 and NRMSE decreased by 1.2%. GPR method provides a methodological reference for selecting crop nitrogen hyper-spectrally sensitive bands and inversion of the nitrogen content of leaves and canopy level during rice different growth period.