Accurate within-field yield estimation is an essential step to conduct yield gap analysis and steer crop management towards more efficient use of resources. This study aims to develop and validate a process-based soybean model and to predict within-field yield variability by coupling leaf area index (LAI) retrieval from Sentinel-2 into the crop model. First, a soybean model is presented, which was successfully validated with field observations of total aboveground biomass, LAI and yield from seven contrasting field campaigns with strongly varying conditions. Within-field yield predictions were achieved by combining the model and the observations of LAI through an assimilation strategy. Four model parameters were chosen to optimize against the LAI curve: soil depth, field capacity, initial LAI and nitrogen translocated from leaves to seed. Six fields were used to evaluate the methodology (21175 pixels). The accuracy assessment was conducted on a pixel-by-pixel basis using high density of information from the yield monitor. The overall accuracy quantified by the relative root mean square error (rRMSE) ranged from 28 to 51% (overall rRMSE 35.8%) across the studied fields. The Lee statistics index ranged from 0.61 to 0.71, confirming a high level of similarity between observed and simulated yield maps. Therefore, the methodology was capable of representing the observed spatial patterns of yield. Furthermore, the high consistency of the optimized WHC reflects the value of the assimilation data strategy to spatialize this relevant characteristic. Some challenges were identified for further study to reduce the sources of uncertainty and improve accuracy: i) the inability of the model to reallocate biomass by simulating plant response to source limitation, ii) the generalization of empirical algorithms to retrieve LAI, and iii) the exploration of an updating method as an assimilation strategy to overcome discrepancy between simulated and retrieved LAI.
Read full abstract