Sugarcane Yield and Field Brix Value Prediction Through Machine Learning Algorithms Using UAV-Based Multispectral Imagery

This study investigates the predictive performance and accuracy of machine learning (ML) algorithms to predict sugarcane (Saccharum officinarum L.) yield and field brix values using unmanned aerial vehicle (UAV)-based multispectral imagery captured at 10th, 11th, and 12th months after planting. UAV imagery was used to calculate five different vegetative indices (VIs) including the normalized difference vegetative index (NDVI), green difference vegetative index (GRVI), excess green vegetation index (ExG), difference vegetation index (DVI), and ratio vegetation index (RVI). Five ML algorithms including multiple linear regression (MLR), partial least square regression (PLSR), random forest (RF), support vector regression (SVR), and extreme gradient boosting (XGB) are used to develop predictive models using VIs. XGB algorithm was able to predict field brix values with a higher coefficient of determination in 10 (R2 = 0.91), 11 (R2 = 0.88), and 12 (R2 = 0.81) months after planting. However, the predictive performances were gradually reduced when reaching closer to the harvesting period (R2 = 10 > 11 > 12 months). When testing each month`s reflectance ability to predict the final yield, same observation was recorded as XGB provided the highest R2 (0.96) values with lower RMSE (0.12) at the 10th month after planting, and interestingly, the predictive performances were also reduced when reaching closer to harvesting period. This result highlights the effects of a weaker correlation of crop canopy reflectance with the agronomic attributes when reaching the harvesting period.