Maize yield estimation from Sentinel-2 multi-temporal imagery and CANbus data integration: a non-parametric regression approach

In precision agriculture, the assessment and estimation of key crop parameters are crucial aspects for the optimisation of input usage and, as an ultimate goal, for the improvement of yield quality and quantity. In this context, a reliable prediction of yield by remotely sensed imagery is an enabling technology for optimisation. In this work, an innovative method for estimating yield in maize cultivation is presented, which exploits multi-temporal and multispectral Sentinel-2 satellite imagery with supervised Machine Learning (ML) techniques. For model training and validation, yield ground truth experimental data from combine harvesters was used, enabling the yield estimation at sub-field scale. The investigation, which was conducted on five case study plots, involved a preliminary comparison of four ML-based algorithms, trained with raw spectral bands. An assessment of the effect of the training dataset on the yield prediction accuracy was then performed. A set of Vegetation Indices (VIs) and Two Band Indices (TBIs) was also considered for this purpose. Finally, a multi-temporal analysis was conducted, in which the temporal evolution of crop spectral data over the maize growing season was exploited using imageries acquired in different epochs. The obtained results proved that an accurate estimation of maize yield can be reached using a Gaussian process regression model, exploiting multi-temporal features directly provided by the raw spectral bands. The model showed a high accuracy in the estimation of maize yield, even when fed with data acquired during only the maize vegetative phase, thus proving its capacity as a prediction tool.