基于无人机数码影像的冬小麦株高和生物量估算

高效、快速地获取作物的株高和生物量信息,对农业生产有重要意义。该文利用2015年4月-6月获得了冬小麦拔节期、挑旗期和开花期的高清数码影像。首先基于无人机高清数码影像生成冬小麦的作物表面模型(crop surface model,CSM),利用CSM提取出冬小麦的株高(Hcsm),然后利用提取的21种数码影像图像指数,构建了拔节期、挑旗期和开花期混合的多生育期生物量估算模型,并进行单生育期和多生育期模型对比分析;最后选择逐步回归(stepwise regression,SWR)、偏最小二乘(partial least square,PLSR)、随机森林(random forest,RF)3种建模方法对多生育期估算模型进行对比,挑选出冬小麦生物量估算的最优模型。结果表明,提取的Hcsm和实测株高(H)具有高度拟合性(R2=0.87,RMSE=6.45 cm,NRMSE=11.48%);与仅用数码影像图像指数构建的生物量估算模型相比(R2=0.721 2,RMSE=0.137 2 kg/m2,NRMSE=26.25%),数码影像图像指数融入H和Hcsm所得模型效果更佳,其中融入Hcsm的模型精度和稳定性(R2=0.819 1,RMSE=0.110 6 kg/m2,NRMSE=21.15%)要优于加入株高H所构建的估算模型(R2=0.794 1,RMSE=0.117 9 kg/m2,NRMSE=22.56%);SWR生物量估算模型(R2=0.7212)效果优于PLSR(R2=0.677 4)和RF(R2=0.657 1)生物量估算模型。该研究为冬小麦生长状况高效、快速监测提供参考。

Estimation of plant height and biomass of winter wheat based on UAV digital image

Abstract: Efficient and timely acquisition of height and biomass of plant is important in improving agricultural management. The purpose of this paper is to investigate the feasibility of using UAV remote sensing to obtain these data. We took winter wheat as an example and conducted a field experiment between April and June 2015 at the Xiaotangshan National Precision Agricultural Research Demonstration Base in Beijing. UAV imageries were taken by a drone from the field at jointing, flagging and flowering stage, respectively. We then developed a crop surface model (CSM) based on these imageries to calculate the plant height and compared the results with field measurements. The image indices extracted from the UAV imageries were used to calculate the biomass using a stepwise regression (SWR) model at each of the three growing stages, as well as the average over the three growing stages. We also compared SWR with the partial least square (PLSR) method and the random forest (RF) method. The results showed that the plant height estimated from the crop surface model agreed well with the measurements with R2=0.87, RMSE=6.45 cm and NRMSE=11.48%. The biomass model was calibrated separately for the jointing, flagging and flowering stage separately, as well as for integrating the three stages as one. Comparison with the measured biomass showed that R2, RMSE and NRMSE of the SWR model were 0.537 4, 0.0500 kg/m2 and 19.13% at the jointing stage, 0.606 6, 0.092 0 kg/m2 and 18.11% at the flagging stage, and 0.6324, 0.117 8 kg/m2 and 14.91% at the flowing stage, respectively. For average biomass over the three stages, R2, RMSE and NRMSE of the SWR model were 0.721 2, 0.137 2 kg·m-2 and 26.25% respectively. It was found that incorporating the plant height into the SWR model improved the biomass estimation, with its associated R2 and NRMSE increasing to 0.794 1 and 22.56% while RMSE reducing to 0.117 9 kg/m2. The SWR model is superior to the PLSR and RF model whose R2 was 0.677 4 and 0.657 10, respectively. In summary, we presented methods to estimate the height and biomass of plant based on UAV imagery and validated it against field experiment with winter wheat as the model plant