Bayesian calibration of AquaCrop model for winter wheat by assimilating UAV multi-spectral images

Crop growth model plays a paramount role in smart farming management, which not only provides quantitative information on crop development but also evaluates various management strategies. A reliable model is desirable but challenging due to the presence of unknown and uncertain parameters; therefore, crop model calibration is significant to achieve its potentials. This work is focused on the calibration of AquaCrop model by leveraging advanced Bayesian inference algorithms and UAV multi-spectral images at field scales. In particular, aerial images with high spatial-temporal resolutions are first applied to obtain Canopy Cover (CC) value by using machine learning based classification. The CC is then assimilated into AquaCrop model and uncertain parameters could be inferred by Markov Chain Monte Carlo (MCMC). Both simulation and experimental validation are performed. The experimental aerial images of winter wheat at Yangling district from Oct/2017 to June/2018 are applied to validate the proposed method against the conventional optimisation based approach by Simulated Annealing (SA). 100 Monte Carlo simulations show that the root mean squared error (RMSE) of Bayesian approach yields a smaller parameter estimation error than optimisation approach. While the experimental results show that: (i) a good wheat/background classification result is obtained for the accurate calculation of CC; (ii) the predicted CC values by Bayesian approach are consistent with measurements by 4-fold cross validation, where the RMSE is 0.0271 smaller than optimisation approach (0.0514); (iii) in addition to parameter estimation, their distribution information is also obtained in the developed Bayesian approach, reflecting the prediction confidence. It is believed that the Bayesian model calibration, although is developed for AquaCrop model, can find a wide range of applications to various simulation models in agriculture and forestry.