Estimating fractional vegetation cover of maize under water stress from UAV multispectral imagery using machine learning algorithms

Abstract Crop water stress is an inevitable and increasing challenge for agriculture. To improve crop water use efficiency, management of water stress and accurate estimation of crop traits were required. Among crop traits used to detect crop growth status and predict yield, fractional vegetation cover (FVC) is of great significance. We conducted studies in a maize field located in Inner Mongolia, China with different irrigation levels during 2018 and 2019 growing seasons. UAV RGB imagery was captured to investigate the effect of image sensors on thresholds obtained by the fixed-threshold method (proposed in our recent study), and to provide reference FVC (FVCUAV_R) for FVC models based on UAV multispectral imagery. Five vegetation indices (VIs), calculated from UAV multispectral imagery, and three regression algorithms (RF: random forest, ANN: artificial neural network, and MLR: multivariate linear regression) were used to build the FVC model suitable for different growing seasons, growth stages, and crop water stress. The results showed that there was a change in thresholds obtained using the fixed-threshold method based on different image sensors, but this change did not make a big difference on the accuracy of FVCUAV_R, with the R2 difference of 0.01 and the RMSE difference of 0.01. As for the three FVC regression models, RF model was the most suitable model when these models established in 2018 were used to estimate maize FVC in 2019 for different growth stages and water stress. The low estimation accuracy for high FVC levels was the reason why MLR model could not be used in the other maize growing season. This study provides a low cost and easy way to estimate maize FVC and its inter-field variability under various water status in different maize growing seasons or growth stages.