Estimation of winter cereals yield exploiting remotely derived phenometrics from LAI time-series

Remote sensing imagery has been extensively used in data-driven approaches to estimate yield employing a variety of approaches. The main goal of this work is to estimate the yield of winter cereals at field level exploiting a multi-year dataset of ground data and satellite derived time series of Leaf Area Index (LAI) with a non-parametric regressive approach. The area of interest is a large farm (~3800 ha) in north Italy where a data set of 183 ground measurements of crop yield (spelt, soft and durum wheat and barley) were collected at the end of 4 cropping seasons (2020-2023). For the same period, Sentinel-2 data were downloaded, and LAI maps were computed using a Gaussian Process Regression (GPR) algorithm already calibrated using a multi-site, crop, and year data set. LAI time-series were exploited to automatically compute phenological metrics (e.g. start and end of season), that are exploited to compute about 20 regressors (e.g cumulated LAI before flowering). These regressors are exploited in a data-driven machine learning approach using multiple random forest (RF) trials. A preliminary test with linear regression shows weak correlations (maximum r2 0.39), while first tests conducted with RF algorithm show that better results can be obtained using all LAI regressors (r2 0.60, RMSE 1.50). Next activities will be first focused on a comparison of LAI estimation between an improved GPR and the ones from Biophysical processor toolbox in SNAP. Phenological metrics will be computed from the more promising source of LAI and RF trials will be performed using the new phenological dates and adding meteorological variables from ERA5 (e.g. rainfall, evapotranspiration) to regressors list. The new “Best model” will be applied on the whole farm and field level estimates and spatial patterns will be analysed and discussed according to farm data.