Monitoring Of Field Maize Growing Under Various Water Stress Conditions Based On UAV Remote Sensing Data

With the continuous growth of the world population,the rapid development of social economy and the continuous changes of the global climate,precision agriculture management has gradually become a current research hotspot.Estimating crop traits,such as field crop coverage,plant height and above-ground biomass,accurately and efficiently contributes to the precision agriculture management.In this study,experiments were conducted across the 2013,2015,2016,2018,and 2019 growing seasons of maize with different irrigation treatments in two sites.Unmanned aerial vehicle(UAV)RGB and multispectral remote sensing systems were used to capture images of field maize at different heights.Spectral reflectance derived from UAV images and ground measurements were used to obtain high resolution FVC map,PH map,and biomass map of maize,providing technological support and instructive guidance for field management in precision agriculture.The main research contents and conclusions of this study were as follows:(1)Estimating maize fractional vegetation cover(FVC)from ground-based RGB imagery.Two classic threshold-based methods,the intersection method(T₁ method)and the equal classification probability method(T₂ method),have been widely applied to RGB images.However,the high crop coverage and severe crop water stress in the field make it difficult to extract FVC stably and accurately.A new method,the fixed-threshold method,was proposed to improve the performance of this FVC estimation.In addition,it was investigated that the influence of different image sensors on the fixed thresholds and whether the change in fixed thresholds,caused by different image sensors,would make a big difference on the accuracy of FVC estimations which were calculated by the fixed thresholds.Results showed that 1)compared with the T₁ and T₂ methods,the mean absolute error(MAE)for the fixed-threshold method reduced by 0.043 and 0.193 respectively,and there is no significant difference in the MAEs of FVC estimation under different water stress conditions;2)the fixed-threshold method was more efficient than the other two methods,with the average time of processing each image reducing by 43.63 and 56.05 s respectively;3)different RGB sensors could cause a certain offset to the initial and the adjusted thresholds calculated based on the fixed-threshold method,but the offset does not make a big difference on the FVC estimation accuracy.(2)Estimating maize fractional vegetation cover map from UAV multispectral imagery.Currently,high-resolution images were required when estimating FVC based on UAV RGB images by using cluster or threshold methods,which means that the reduction of image resolution will significantly reduce the FVC estimation accuracy.However,there are few studies about estimating crop FVC map based on the combination of UAV multispectral imagery and machine learning methods which requires a lower resolution for images.It was deserved to further investigate the suitability of the regression models to different growth stages,different growing seasons,and different water stress conditions.Therefore,in this study,random forest(RF),artificial neural network(ANN),and multi-linear regression methods were adopted to build FVC regression models based on five vegetation indices(VIs)related with FVC.In addition,the optimal model was selected by comparing the validation performance of the three models to obtain the FVC map of maize.Results showed that 1)compared to the ANN and MLR regression models(MAE of 0.12 and 0.23),the RF regression model performed best with the MAE of 0.05.For three different water stress conditions,the RF regression model performed best with the R²adj of 0.94,0.90,and 0.86,MAE of 0.03,0.06,and 0.05.For different growth stages,the RF performed best with MAE of 0.05,0.05,and0.04.All results indicated that the RF regression model had the best suitability,followed with the ANN regression model.The MLR regression model had the worst suitability,which may be caused by the unstable performance when the MLR regression model was applied to the maize with high coverage.(3)Estimating maize plant height based on crop surface model constructed from UAV RGB images.At present,the most commonly used method of crop height estimation based on remote sensing is to extract the height features from the crop surface model CSM.In this method,the digital ground model DTM is a key parameter.However,to the best of our knowledge,no study has been reported to investigate under what FVC conditions DTM could be constructed accurately based on the selected ground elevations and interpolation algorithms.In addition,the influence of oblique and nadir view and spatial resolution on maize PH estimation accuracy in a large farm still needs further study.In this study,the experiment was conducted in a maize field.The influence of fractional vegetation cover(FVC)on DTM construction accuracy was investigated for the first time,and the influences of view angle(oblique and nadir)and spatial resolution on the accuracy of maize PH estimation were explored.This study provided an instructive guidance for the optimal PH estimation method.Results showed that 1)compared to the DTM constructed using UAV images over bare soil shortly after sowing,FVC less than 0.4 was necessary for accurate construction of DTM,with averaged estimation errors of 15 cm in 2018 and 9 cm in 2019;2)the optimal PH feature,with the smallest errors compared to ground-truth PH,was the 99^th percentile in 2018(root mean square error(RMSE)of 22.69 cm,mean absolute error(MAE)of 18.64 cm)and 100^thpercentile in 2019(RMSE of 18.73 cm,MAE of 14.53 cm),indicating that compared to the nadir view,the oblique view resulted in a more accurate 3-D reconstruction;3)when the original spatial resolution of 15 mm was upscaled to 20,30,60,and 120 mm by the bilinear interpolation algorithm,a decreasing trend of PH estimation accuracy was observed,with RMSE increasing from 34.70 to 39.98 cm and MAE increasing from 28.96 to 36.15 cm.(4)Estimating maize biomass based on multi-source remote sensing.Currently,although the UAV remote sensing system was widely applied to the crop biomass estimation,it is difficult to construct a high-precision biomass estimation model suitable for the whole growing season.Because the spectral and structural parameters obtained from UAV remote sensing saturated for high biomass.In this study,the experiment was conducted in a maize field across the 2018 and 2019 growing seasons.The combination of growing-degree days(GDD)and vegetation indices,FVC_UAV,PH_UAV derived from UAV remote sensing was used to construct a biomass estimation model suitable for maize at vegetative,reproductive,and maturity growth stages.In addition,a biomass regression model suitable for different growing seasons was determined by comparing the applicability of the MLR and RF regression models in different growing seasons.Results showed that 1)the combination of GDD and VIs solved the saturation problem that happens at reproductive and maturity growth stages;2)compared to the biomass regression model based on only VIs,the R² obtained from the model based on the combination of GDD and VIs was increased by 0.06?0.09 for fresh biomass estimation,and 0.07?0.23 for dry biomass estimation,r RMSE was decreased by 0.02 for fresh biomass estimation,and 0.03?0.07 for dry biomass estimation;3)there is no significant difference between MLR and RF regression models when the combination of GDD and VIs were used.However,the MLR regression model built based on the dataset from the 2018 growing season failed in applying to the dataset from the 2019 growing season,and the RF regression model had a relatively better performance with R² of 0.59 and r RMSE of 0.19.Overall,this study improved the biomass estimation accuracy at reproductive and maturity growth stages and provided a way of constructing regression models suitable for different growing seasons.