Estimating spatial and temporal variability of crop growth by radiation-driven models based on satellite data assimilation

Precision agriculture aims to improve field management by accounting for spatial variability in dynamic cropping systems (CS). To support farmers in optimizing crop management, it has become important to develop tools that capture both spatial and temporal heterogeneity within crop production. Mechanistic crop models simulate aboveground biomass (AGB) accumulation by explicitly representing physiological and environmental processes linking absorbed solar radiation, transpiration, and nutrient uptake. However, these process-based models typically require site- and cultivar-specific calibration and often fail to capture fine-scale variability. Assimilating remote sensing (RS) data into radiation-driven Production Efficiency Models (PEM; Monteith, 1977; McCallum et al., 2009), offers a promising solution by providing spatio-temporal explicit predictions of biomass growth, and thus crop requirements or responses. A PEM was developed and implemented in a PostGIS database to run at a daily time steps and with Sentinel-2 (S2) resolution (10 x 10 m). Leaf Area Index (LAI), derived from the S2 biophysical processor (Weiss et al., 2020), was assimilated daily to estimate the fraction of intercepted solar radiation by crops at pixel level via the Lambert–Beer law. The PEM was calibrated and evaluated for wheat and maize grown in the Po Valley (northern Italy) using two independent datasets. Calibration relied on 120 AGB observations (2022-2023) from five different sites with different management practices, while evaluation was carried out on 312 observations (2025) from four additional sites. Although the temporal and spatial dynamics of LAI should implicitly reflect the effects of limiting factors, a development stage-dependent Morris sensitivity analysis (SA) was conducted to assess how variations in LAI dynamics, air temperature, and senescence affect simulated AGB. The simulations performed for each S2 pixel provide reliable estimates of AGB. On evaluation, the model achieved a relative Root Mean Square Error of 0.23 and a model efficiency of 0.92, effectively capturing temporal and spatial variability across and within the different CS (Figure 1). Assimilating LAI enabled the model to overcome common limits of process-based models that require site- and crop-specific tuning. SA highlighted that LAI accuracy during early growth stages strongly affects AGB predictions and in-season decisions. From stem elongation onward, radiation use efficiency (RUE) and temperature responses became dominant drivers, with senescence adjustment from milk ripening stage preventing overestimation at harvest. Integrating daily LAI assimilation into a simplified radiation-driven framework captured spatial and temporal variability across crops, sites, management systems and seasons. Accurate early-season LAI estimates are critical, while RUE, temperature, and senescence primarily drive AGB growth in later stages. By assimilating RS data, the model overcomes common limitations of process-based approaches and provides a robust tool for monitoring crop productivity across heterogeneous CS.