Regional Model to Predict Sugarcane Yield Using Sentinel-2 Imagery in São Paulo State, Brazil

Sugarcane yield prediction is an important tool to support the sugar-energy sector. This study aimed to create a regional empirical model, using the random forest algorithm, to predict sugarcane yield in the state of Sao Paulo. For this, we used Sentinel-2 imagery (vegetation indices NDVIRE and CIRE, spectral bands Red-edge and near-infrared arrow), agronomic data (variety and ratoon stage and plant cane), climatic data (temperature, precipitation) and crop water deficit data from three mills. We created two predictive yield model based on three scenarios with different training and testing data: (SI) Scenario I is the regional model considered all data from the three mills, (SII) Scenario II was training similar SI and testing individuals for each mill, (SIII) Scenario III includes regional individual’s models for sugarcane ratoon stage and plant cane. In each case, 70% of the dataset was used for training and 30% for testing. SI gave R² equal to 0.72, while SII R² was between 0.60 and 0.78; the RMSE for SI was 11.7 \({\text{tonha}}^{{ - 1}}\) ${\text{tonha}}^{-1}$ , while for SII from 8.62 to 15.56 \({\text{tonha}}^{{ - 1}}\) ${\text{tonha}}^{-1}$ . The rRMSE was 16.5% for SI and from 12.4 to 21.6%, for SII. SIII showed R² greater than 0.61, and RMSE between 9.6 and 13.5 \(ton {ha}^{-1}\) $ton{\mathit{ha}}^{-1}$ . The CIRE and NDVIRE vegetation indices, crop water deficit and precipitation were the most important variables to estimate sugarcane yield. The model created considering SI and SII showed potential to be applied to different locals using data from three mills.