基于无人机高光谱融合连续投影算法估算棉花地上部生物量

Estimation of cotton above-ground biomass based on unmanned aerial vehicle hyperspectral and successive projections algorithm

【目的】地上部生物量是表征植物生命活动的重要参数。探索不同的光谱预处理方法和建模方法,实现对棉花地上部生物量快速、无损、准确的估算,对棉花长势监测和大田精准管理具有重要意义。【方法】以新陆早53号、新陆早45号为研究对象,设置不同施氮处理,于出苗后不同阶段获取棉花地上部生物量和无人机高光谱数据,通过连续投影算法(Successive projections algorithm,SPA)筛选不同预处理[一阶导数、二阶导数、Savitzky-Golay(SG平滑)、多元散射校正]后的特征波长,基于筛选出的不同波长组合使用偏最小二乘法回归(Partial least square regression,PLSR)和随机森林回归(Random forest regression,RFR)分别构建棉花地上部生物量估算模型,比较不同预处理后建立模型的精度,确定最优估算模型。【结果】(1)利用SPA算法对不同预处理后的光谱信息筛选出特征波长9~26个,可实现光谱信息降维。(2)基于SG平滑-SPA处理及PLSR方法建立的模型最佳,R^(2)达到了0.63,均方根误差(Root mean square error,RMSE)为0.42,验证集的R^(2)为0.67,RMSE为0.44。(3)一阶导数-SPA处理后,采用RFR构建的模型最佳,R^(2)达到0.87,RMSE为0.45,验证集R^(2)为0.81,RMSE为0.37。【结论】采用一阶导数预处理结合SPA筛选特征波长,经RFR构建的估算模型结果和验证效果均最佳,可用于棉花地上部生物量定量估算。

[Objective]Above-ground biomass is an important parameter of plant life activity.Exploring different spectral pretreatment methods and modeling methods to achieve rapid,nondestructive and accurate estimation of cotton above-ground biomass is of great significance for cotton growth monitoring and field precision management.[Method]Xinluzao 53 and Xinluzao 45 were selected as the research objects,and different nitrogen application treatments were set up to obtain cotton above-ground biomass and UAV hyperspectral data at different stages after emergence.The successive projections algorithm(SPA)was used to select the characteristic wavelengths after different pretreatments[first derivative,second derivative,Savitzky-Golay(SG smoothing),multiple scatter correction].Based on the selected wavelengths,partial least squares regression(PLSR)and random forest regression(RFR)were used to construct cotton above-ground biomass(dry matter)estimation models,respectively,and the optimal estimation model was determined by comparison of the model estimation accuracy.[Result](1)The SPA algorithm can effectively remove the redundant bands from the spectra,and the number of above-ground biomass-sensitive characteristic wavelengths after different pretreatments ranges from 9 to 26,which could be used to reduce the dimension of spectral information.(2)Based on SG smoothing-SPA processing,the model established by PLSR is the best,with a coefficient of determination(R^(2))of 0.63,root mean square error(RMSE)of 0.42,and validation set R^(2),RMSE of 0.67,0.44.(3)After the first derivative-SPA processing,the model constructed by RFR is the best,with R^(2)of 0.87,RMSE of 0.45,validation set R^(2)and RMSE of 0.81 and 0.37.[Conclusion]Using the first derivative pretreatment combined with SPA to select biomass-sensitive wavelengths,the RFR model has the best results and validation performance,which can be used for quantitative estimation of cotton above-ground biomass.