基于无人机影像的冠层光谱和结构特征监测甜菜长势

甜菜是中国北方地区重要的经济作物。快速、准确、高通量的获取甜菜的地上部和块根鲜质量、块根含糖率、叶绿素含量对甜菜生产具有重要意义。该研究采用无人机搭载数码和多光谱相机，获取甜菜叶丛快速生长期、块根及糖分增长期和糖分积累期的数码影像和多光谱影像，提取了冠层的结构特征和光谱特征。选择随机森林回归（Random Forest Regression，RFR）和偏最小二乘回归（Partial Least Squares Regression，PLSR）2种建模方法基于获取的冠层特征，构建甜菜全生育期的地上部和块根鲜质量、块根含糖率和SPAD（Soil and Plant Analyzer Development）值估算模型。研究结果表明，随机森林回归模型和偏最小二乘回归模型对地上部和块根鲜质量、含糖率都做出较好的预测，R2范围分别为0.9～0.94、0.88～0.9，rRMSE范围分别为7.6%～17%、8.8%～20%。对SPAD值的预测均较弱，R2分别为0.66和0.67。为了减小输入变量集的大小以及去掉对预测不敏感的变量，该研究采用置换重要性（Permutation Importance，PIMP）来筛选冠层光谱特征和结构特征中对预测有重要影响的变量。结果表明基于筛选出的重要性特征构建的随机森林回归模型和偏最小二乘回归模型对地上部和块根鲜质量、含糖率都做出较好的预测，R2范围分别为0.89～0.94、0.74～0.91，rRMSE范围分别为7.3%～19%、7.6%～19%。对SPAD值的预测均较弱，R2分别为0.65和0.68。进一步表明随机森林回归模型在精度上略好于偏最小二乘回归模型。同时基于PIMP筛选变量的方法在保持原有精度的同时能实现降低数据收集复杂性的目的。研究结果为基于无人机遥感技术快速、准确监测甜菜长势和估测块根类作物的根部活性物质提供了参考。

Monitoring of sugar beet growth using canopy spectrum and structural characteristics with UAV images

Abstract: A sugar beet is one of the most important cash crops in northern China. It is a high demand for the rapid, accurate, and high-throughput acquisition of the fresh weight of aboveground and root, the sugar content of root, and the chlorophyll content of aboveground in the production of sugar beet. An Unmanned Aerial Vehicle (UAV) can serve as a significant approach, due to its flexibility, low cost, and high spatiotemporal resolution. In this study, a UAV equipped with digital and multispectral cameras was utilized to capture the images of sugar beet during the leaf clusters, root tuber, sugar growth, and accumulation period, thereby extracting the structural and spectral characteristics of the canopy. The estimation models were also established for the various indexes using the Random Forest Regression (RFR) and Partial Least Squares Regression (PLSR), including the fresh weight of shoot and root tuber, the sugar content of root tuber, and Soil Plant Analysis Development (SPAD) value during the whole period of sugar beet. The results showed that the RFR and PLSR model performed well to predict the fresh weight and sugar content of shoot and root tuber, with the coefficient of determination R2 ranging from 0.9 to 0.94 and from 0.88 to 0.9, respectively, while the relative Root Mean Square Error (rRMSE) ranging from 7.6% to 17% and from 8.8% to 20%, respectively. Both models presented weak predictions for the SPAD values, where the R2 values were only 0.66 and 0.67, respectively. Furthermore, a Permutation Importance (PIMP) was used to screen the more sensitive variables with the dominated impacts on the prediction, in order to reduce the size of the input variable set for the less cost and complexity of data collection. As such, the optimal prediction models of RFR and PLSR were achieved for the growth monitoring of sugar roots. It was found that excellent predictions were achieved on the fresh weight and sugar content of shoot and root tuber, with the R2 value ranging from 0.89 to 0.94, and from 0.74 to 0.91, respectively, and the rRMSE value ranging from 7.3% to 19% and from 7.6% to 19%, respectively. Nevertheless, the RFR and PLSR model presented weak predictions for the SPAD values, where the R2 values were only 0.65 and 0.68, respectively. Correspondingly, the accuracy of the RFR model was slightly better than that of the PLSR model. More importantly, the PIMP variable screening can be widely expected to reduce the complexity of data collection with optimal accuracy. Consequently, the canopy structure and spectral features obtained by UAVs can be utilized to quickly and accurately monitor the growth and sugar content of sugar beet. The finding can provide a strong reference to estimate the root active substances of tubers crops using UAV proximity.