基于无人机多光谱影像的玉米涝渍胁迫产量估算方法

Estimation of Maize Yield under Waterlogging Stress Based on UAV Multispectral Images

摘要:

涝渍胁迫是玉米生产的主要气象灾害之一,快速准确的涝渍胁迫产量估算对灾害管理与田间生产决策至关重要。传统田间人工观测方式难以满足突发性作物灾害评估的时效性要求。本研究基于水深、历时和品种的多变量拔节期玉米涝渍模拟试验,获取了3个时期的无人机多光谱影像,采用皮尔逊相关系数法筛选与产量敏感的植被指数,使用随机森林算法构建玉米产量估算模型,探究无人机多光谱影像估算涝渍胁迫下玉米产量的应用效果。研究结果表明:耐涝品种MY73表现出最优的抗逆性与高产潜力,不同处理下的MY73产量为4300~10500kg/hm2,而所有品种在淹水及胁迫时间延长条件下产量均显著下降。不同植被指数对于产量的敏感性随着无人机影像获取时间不同而存在差异,其中任意时间点获取的EVI、DVI 均与产量显著相关。随机森林算法与敏感植被指数相结合可实现涝渍胁迫下的玉米产量高精度估算,多时期协同的玉米产量估算模型的稳定性优于单一时间点,测试集R2为0.86,RMSE为954.15kg/hm2,MAE为700.65kg/hm2,NRMSE为10.50%。因此,无人机多光谱遥感技术可以实现涝渍胁迫下的玉米产量快速估算。

Abstract:

Waterlogging stress is one of the major meteorological disasters in maize production. Rapid and accurate estimation of maize yield under waterlogging stress is crucial for disaster management and field production decision-making. Traditional field-based manual observation methods are difficult to meet the timeliness requirements of disaster assessment. Multi-variable waterlogging simulation experiments were conducted during the maize jointing stage, incorporating factors such as water depth, submergence duration, and cultivar variation. Three sets of unmanned aerial vehicle (UAV) multispectral images were acquired throughout the experimental period. Vegetation indices highly correlated with yield were identified by using the Pearson correlation coefficient method, and a maize yield estimation model was developed based on the random forest algorithm. This study aiming to evaluate the potential of UAV-based multispectral remote sensing for estimating maize yield under waterlogging stress. The results showed that the waterlogging-tolerant variety MY73 exhibited the strongest stress resistance and the highest yield potential. The yield of MY73 under different treatments ranged from 4300kg/hm2 to 10500 kg/hm2. However, the yield of all varieties significantly decreased with the increase of flooding severity and duration. The sensitivity of vegetation indices to yield varied with the acquisition time of UAV images, among which EVI and DVI were significantly correlated with yield at all time points. The combination of the random forest algorithm and sensitive vegetation indices enabled high-accuracy estimation of maize yield under waterlogging stress. The stability of the multi-period collaborative model for estimating maize yield per unit area was superior to that of a single time point, with an R2 of 0.86, RMSE of 954.15kg/hm2, and MAE of 700.65kg/hm2, and NRMSE of 10.50% on the test set. Therefore, unmanned aerial vehicle multispectral remote sensing technology demonstrated the capability to achieve rapid estimation of maize yield under waterlogging stress.