A machine learning models approach and remote sensing to forecast yield in corn with based cumulative growth degree days

Pre-harvest yield forecasting is important for the sustainability of agricultural companies, enabling more sustainable economic decision-making. In the present study, we propose an approach based on the sum of degree days of the corn crop related to the dates of satellite images to organize the data of two crops generating predictive models with the k-nearest neighbors (KNN) and extreme gradient boosting (XGBoost). The field study was carried out in a commercial area during the 2017/18 and 2018/19 harvests. Spectral data were obtained from Sentinel-2 satellite images. After the correction and processing of the images, the values ​​of the spectral bands and the vegetation indices were obtained. For the development of the models, the images obtained throughout the cycle were divided into three classes of the mean weeks before harvest (WBH) from different degree-days (GD) during the cycle, in this study we adopted 12 combinations of data inputs to develop the models. In yield forecasting, we were able to forecast approximately 30 to 70 days before harvest (500 to 900 degree-days before harvest), in addition, the most accurate models were when the data used as driven variables were the spectral bands of the red, blue, green and nir collected from 800 to 1200 degree-days of the culture (WBH4). For the models developed, combined with WBH for yield forecast, it was possible to forecast yield with an average error of 0.503 t ha-1, and the greatest precision and accuracy occurred with the use of all variables RGB e Near-infrared.