Assimilated Remote and Proximal Sensing Improved Crop Model Wheat Yield Prediction Accuracy

Abstract

This project tests if assimilating remote and proximal sensing observations into the APSIM crop model can improve wheat yield prediction accuracy without calibrating against observed yield. This is achieved by using a sensitivity analysis to map the weighting from the errors between crop condition simulations and observations with the impact of individual model parameters on outputs and taking the mean of an ensemble of predictions. This thesis consists of chapters on: (1) a review on relevant literature, (2) remote sensing of field (wheat) experiments over the 2019 and 2020 seasons, (3) sensitivity analysis of the APSIM crop model; and, (4) Data Assimilation (DA) framework and performance. Results indicate that remote sensing has strong potential for sampling crop status with high spatial, spectral and temporal resolution, as seen with accurate protein concentration estimation (r2 of 0.88 in 2020) and strong yield prediction (r2 of 0.64 in 2019). There was limited consistency between seasons, however. ‘Sobol’ main effect indices were calculated for 23 parameters in APSIM and averaged across the season and then used in the crop modelling DA framework. In the DA, ensemble members covering 4 key areas of uncertainty were run over 15 iterations of the 2019 and 2020 seasons, with their parameters modified between iterations using the relative errors in their predictions scaled by the sensitivity indices. Error weighting was derived from comparing predicted height growth, water extraction and yield against height growth from RGB photogrammetry, water extraction from soil moisture probes and potential yield from a statistical model. The mean of each ensemble was tested against observed yield for the year as well as APSIM performance without DA. Results indicate there is strong potential for this approach, with average accuracy improvements between 7-12% of observed yield in all tests.