To quickly and accurately obtain the moisture status of grape plants at the field scale, the treatments included three irrigation levels i.e. W3 (100%M, M as the irrigation quota), W2 (75% M) and W1 (50%M) and four fertilizer application rates i.e. F0 (0 kg/hm²), F1 (324 kg/hm²), F2 (486 kg/hm²) and F3 (648 kg/hm²). Grape leaf water content (LWC) was monitored nondestructively by an unmanned aerial vehicle (UAV) carrying multispectral (MS), visible light (RGB) and thermal infrared (TIR) cameras to extract band reflectance (BR), canopy coverage (CC) and canopy temperature (T) information, respectively. Using BR (included six bands: B, R, G, RE, NIR₈₀₀, and NIR₉₀₀), CC and T and their combinations as input variables brought into the partial least squares (PLS), random forest (RF), support vector machine (SVM) and extreme learning machine (ELM) algorithms. The prediction models for grape LWC were established by using four machine learning algorithms, and the optimal combination of variables was finally determined. The results represented that (1) the model built with BR + CC + T as predictor variables under different water treatments was better than other combinations of variables, with the coefficient of determination (R²) above 0.69 and the relative root mean square error (RRMSE) less than 2.5%; (2) modeling the LWC of grapes at different fertility periods based on the combination of BR + CC + T, the R² ranged from 0.51 to 0.78 at the shoot-growing, anthesis, and fruit-inflating stages; (3) the top three important variables were T, NIR₈₀₀, and NIR₉₀₀ in the shoot-growing, anthesis, and fruit-inflating stages, while the top three important variables were RE, B, and T in the fruit-inflating stage. In summary, UAV multimodal data fusion has good application in predicting the LWC of grapes using RF algorithm modeling during the different growth stages. This study can supply a technical support for precise management of vineyard water regime using a UAV platform.

为了快速、准确地获取大田尺度葡萄植株的水分状况，处理包括W3（100%M，M为灌溉定额）、W2（75%M）和W1（50%M）三个灌溉水平，四种施肥量，即F0（0 kg/hm ²）、F1（324 kg/hm ²）、F2（486 kg/hm ²）和F3（648 kg/hm ²）。通过携带多光谱 (MS)、可见光 (RGB) 和热红外 (TIR) 相机的无人机 (UAV) 对葡萄叶水分含量 (LWC) 进行无损监测，以提取波段反射率 (BR)、冠层覆盖度 (CC) 和分别是冠层温度（T）信息。使用BR（包括6个波段：B、R、G、RE、NIR ₈₀₀和NIR ₉₀₀）、CC和T及其组合作为输入变量带入偏最小二乘法（PLS）、随机森林（RF）、支持向量机器（SVM）和极限学习机（ELM）算法。利用四种机器学习算法建立了葡萄LWC的预测模型，最终确定了最优的变量组合。结果表明：（1）不同水处理下以BR+CC+T为预测变量建立的模型优于其他变量组合，决定系数（R ²）在0.69以上，相对均方根误差（ RRMSE）小于2.5%； (2)基于BR+CC+T组合对葡萄不同生育期的LWC进行建模，芽生期、开花期、膨果期R ²范围为0.51～0.78。 (3)芽生期、开花期、膨大期重要变量前三位为T、NIR ₈₀₀、NIR ₉₀₀ ，膨果期重要变量前三位为RE、B、T 。综上所述，无人机多模态数据融合在利用RF算法建模预测葡萄不同生长阶段的LWC方面具有良好的应用前景。该研究可为无人机平台精准管理葡萄园水情提供技术支撑。