Co-design of crop monitoring-modeling systems via identifiability analysis

Crop yield increases are lagging those required to avoid major mid-century food security disruptions. Skilled models are needed to predict environment-specific cropping outcomes for applications ranging from accelerating breeding program gain rates via efficient selection to sustainable, in-field production decision-making that incorporate genetic × environment × management interactions. In an attractive vision, genotype-specific ecophysiological crop model parameters are inferred by genomic prediction. However, these parameters are often unidentifiable from typically observed data, also known as the equifinality problem. This is because wide parameter value ranges generate identical predictions so alternative estimates are indistinguishable and, thus, unusable in genomic prediction training. Given a model with parameters, a prescribed data scheme and goodness-of-fit objective function this paper introduces a method to measure identifiability and semi-automate its improvement. Identifiability is innovatively quantified as the objective function’s Gaussian curvature at the parameter optimum. Higher values indicate a more easily searchable, bowl-like objective function response surface. Lower values imply a flatness that prevents unique parameter value identification. A simplified, seven-parameter wheat model is analyzed, revealing two nearly unidentifiable parameters. Three analyses are presented: (1) increasing observation numbers, (2) adding a new observation type, and (3) automatically determining two parameters whose direct field measurement would maximally increase Gaussian curvature. Two nearly unidentifiable parameters – Thermal time to end of leaf expansion (TTL) and thermal time to maturity (TTM) – were revealed. Analyses (2) and (3) drastically improved identifiability. After adding the new observation type (physiological maturity), Gaussian curvature increased by a factor of 7600. Likewise, when we fixed the values of the two parameters, the Gaussian curvature increased by a factor of 2 . 69 × 1 0 11 . Surprisingly, however, increasing observation numbers lowered identifiability for reasons related to the Nash–Sutcliffe Efficiency objective function variant used. Relatedly, because the method is formally objective function agnostic, it can also ameliorate any non-uniqueness in model-based production decision aids that optimize economic objectives while integrating a full suite of in-field interactions. The demonstrated ability to maximize identifiability by optimizing both the data taken and model structure prior to actual fieldwork points toward a co-design process in which the model and its supporting monitoring system are created simultaneously. • Crop model predictions can aid crucial tasks from breeding to in-field management • Crop model parameter unidentifiability can go unnoticed during calibration • We score identifiability as the Gaussian curvature (GC) of the objective function • Steps that raise GC of the model/inputs/objective function improve identifiability • Our objective-function-agnostic method enables modeling/monitoring systems codesign