On Bayesian inference and remote sensing imagery of state and parameter estimation for crop model

Inefficient use of water resources induces a surge of interest in the improvement of irrigation management to optimally utilise water resources. Researches have shown that a number of technologies have been widely applied to this end, such as remote sensing technologies, crop models, and data assimilation methods. Crop growth models play a paramount role, which not only provide quantitative information on crop development but also evaluate various management strategies. Thus, field-specific crop models are required for reliable and efficient crop treatment and monitoring. Due to the presence of unknown and uncertain parameters in crop models, however, crop state and parameter estimation is generally a challenging problem. Current literature on this challenge is mainly based on either optimisation method to calibrate crop models or conventional filtering method to recursively infer the crop states by using sensors of low spatial/spectral resolutions.

This thesis aims to develop efficient and accurate crop state and parameter estimation methods by using remote sensing technologies and advanced Bayesian inference algorithms. The Bayesian inference framework was firstly employed in state and parameter estimation of crop models, which was achieved through Markov Chain Monte Carlo (MCMC) and sensitivity informed Particle Filter (PF). Different from optimisation-based approaches where only a point estimate is returned without confidence information, Bayesian estimation could derive the posterior distribution of model parameters of interest by integrating the prior information and measurement information via Bayesian rule. The proposed PF was able to accommodate extra information (e.g., parameter bound information and parameter sensitivity information) for better estimation performance. Two state-of-the-art sensors, including MSI on Sentinel-2A and RedEdge camera on UAVs, were analysed by using the Support Vector Machines (SVMs) and the Random Forest classifier to support observation data collection. All proposed estimation methods were successfully implemented, tested and validated by using both Monte Carlo simulations and real-world datasets. Winter wheat was selected as one case study, and the results suggested that the winter wheat state and parameter estimation employing Bayesian-based method outperforms conventional approaches (e.g., optimisation-based estimation and conventional PF). As a consequence, with the advances of the new sensors and the proposed Bayesian inference, there is a major step to reduce model uncertainties in a promising and powerful way such that the accurate irrigation management can be made before harvest.