Precise and reliable yield forecast tools could play a fundamental role in supporting policy formulation,

and decision-making process in agriculture (e.g. storage and transport). Most models developed for yield forecasting are only useful at large but as agricultural practices are oriented towards more site-specific, moving from larger scales to precision agriculture (PA) techniques, there is a higher dependency on detailed information about within-field variability scales. Therefore, to estimate corn (Zea mays L.) yields at the field level is of great interest to farmers, service dealers, researchers, and policy-makers. The main objectives of this study were to: i) provide guidelines on data selection aimed at building forecasting yield models using Sentinel-2 satellite imagery; ii) compare different statistical approaches (Ordinary least square – OLS multivariable regression considering errors independent and identically distributed (i.i.d.) versus models considering errors spatially correlated) and different vegetation indices (VIs) during model building; and iii) perform spatial and temporal validation to see if empirical models could be applied to other regions or when models coefficients should be updated. Data analysis was divided into four major steps: i) data acquisition and preparation; ii) selection of training data; iii) building of forecasting yield models; and iv) spatial and temporal validation. The analysis were performed using yield data collected from 19 corn fields located in Brazil – Rio Grande do Sul state (2016/2017 season) and Mato Grosso state (2016 and 2017 seasons) – and in the United States – state of Kansas (2016 season), and VIs (NDVI, green NDVI and red edge NDVI) derived from Sentinel-2 satellite imagery. All the analysis involving imagery processing were performed in Google Earth Engine, and the analysis involving tabular data and statical analysis were performed in R (R Core Team, 2017). Main outcomes from this study were: i) data selection impacted yield forecast model and fields with narrow yield variability and/or with skewed data distribution should be avoided because they result in more restrictive models (low applicability along the time and space); ii) models that assumed errors being spatially correlated overcome the standard ones considering i.i.d.; iii) red edge NDVI was most retained VI for among all the fields; and iv) model prediction power was more sensitive to similarities on yield data frequency distribution (position of the mode and interquartile range) than to the geographical distance or years. Thus, this study provided guidelines to build more accurate corn yield forecasting models, but also established limitations for up-scaling, from farm-level to county, district, and state-scales.