基于无人机数码影像的马铃薯生物量估算

株高和植被覆盖度（Vegetation Coverage, VC）是估算生物量的重要参数，而生物量的准确估算对农业生产具有重要作用。该研究获取马铃薯现蕾期、块茎形成期、块茎增长期、淀粉积累期和成熟期的无人机和地面数码影像，并实测株高、地上生物量和地面控制点（Ground Control Point, GCP）的三维空间坐标。首先基于数字表面模型（Digital Surface Model, DSM）提取马铃薯株高，其次利用地面和无人机数码影像提取马铃薯VC实测值和估测值，然后将提取的株高、VC和二者乘积与选取的11种植被指数和生物量作相关性分析，挑选出相关性较好的前6种植被指数和3种农学参数，最后通过线性回归（Linear Regression，LR）、偏最小二乘回归（Partial Least Square Regression，PLSR）、随机森林（Random Forest，RF）算法和支持向量机（Support Vector Machine, SVM）估算生物量。结果表明，提取株高和实测株高拟合的决定系数为0.86，标准均方根误差为13.42%；提取VC值和实测VC值拟合的决定系数为0.84，标准均方根误差为15.76%；利用LR建模和验证精度由低到高依次为提取的株高、VC和二者乘积，每种变量的估算效果均从现蕾期到块茎增长期逐渐变好，从淀粉积累期到成熟期逐渐变差；每个生育期利用3种方法以不同变量估算生物量效果依次由低到高为植被指数、植被指数结合提取株高、植被指数结合提取VC、植被指数结合提取的株高和VC，其中PLSR模型效果优于RF和SVM模型。该研究为马铃薯长势快速监测提供参考。

Estimation of potato biomass based on UAV digital images

Accurate estimation of crop biomass by plant height and Vegetation Coverage (VC) is of great significance in agricultural production and has a strong guiding significance for agricultural managers. It is necessary to use an effective method to estimate the biomass of field crops quickly and accurately. Taking the potato in Xiaotangshan National Precision Agricultural Research Demonstration Base as the research object and conducted a field experiment between March and July 2019. Digital images were taken by Unmanned Aerial Vehicle (UAV) and ground camera from the field at budding, tuber formation, tuber growth, starch accumulation, and maturity period and measured plant height, above-ground biomass, and three-dimensional coordinates of ground control points (GCPs) were obtained by ground survey. Firstly, the Digital Surface Model (DSM) is generated based on GCPs to calculate the plant height and compared the results with field measurements. Secondly, Measured and estimated values of VC were calculated through ground and UAV digital images and compared between the results. Then, the correlation analyses between biomass and extraction values of plant height, VC, their product, and eleven Vegetation Indices (VIs). Six VIs and three agronomy parameters were selected for each growth stage, respectively. Finally, the selected VIs and three agronomy parameters were used as modeling factors, and the biomass was estimated by Linear Regression (LR), Partial Least Square Regression (PLSR), Random Forest (RF) algorithm, and Support Vector Machine (SVM), and the models constructed by the remote sensing data were compared to optimize the model. The results showed that the extraction values of plant heigh from DSM agreed well with the measurements, the coefficient of determination was 0.86 and the normalized root mean square error was 13.42% throughout the growth period. Measured and predicted values of VC stayed highly relevant, the coefficient of determination was 0.84 and the normalized root mean square error was 15.76% throughout the growth period. Through three agronomy parameters, analyzing the effect of the modeling and verification set, the accuracy of the Linear Regression (LR) model with extraction values of plant heigh multiply predicted values of VC as modeling factors was significantly better than that of extraction values of plant heigh or predicted values of VC, however, the accuracy of estimating biomass model with extraction values of plant height was the worst. In different growth periods, the performance of estimating biomass by LR had gradually increased from the budding stage to the tuber growth stage and reduced from the starch accumulation stage to the maturity stage. To compare capabilities of PLSR, RF, and SVM to estimate potato biomass, this study compared the accuracy of models for different growth periods using four variables, for example, VIs combined with extraction values of plant height, VIs combined with extraction values of VC and three variables as one. For PLSR, RF, and SVM models, the accuracy of modeling and verification showed a trend of first increasing and then decreasing when using the same kind of variables as model factors. Comparison of the accuracy of models was contrasted by different methods with the same variable at five periods, it is found that VIs incorporating the plant height and VC into estimation model significantly improved the biomass estimation. Comparison with the measured biomass showed that the coefficient of determination, the normalized root mean square error of PLSR model was 0.628 5 and 18.21% at bud period, 0.658 4 and 15.41% at tuber formation period, 0.681 4 and 11.42% at tuber growth period, 0.653 2 and 20.61% at starch accumulation period, 0.548 8 and 21.34% at maturity period, respectively. The PLSR model is superior to the RF and SVM model which the coefficient of determination was 0.538 5 0.603 3, 0.632 2, 0.615 9, 0.542 4 and 0.445 1, 0.521 1, 0.601 3, 0.574 3, 0.538 4, respectively. In summary, the biomass of potato was quickly estimated using UAV digital images data by the PLSR method combined with VIs, plant height, and VC as a whole in different growth periods and provided technical support for effectively monitoring crop growth and accurately predicting yield.