Field-level sugarcane yield estimation utilizing Sentinel-2 time-series and machine learning

This work focused on developing a methodology for using machine learning (ML) approaches to establish a pre-harvest yield prediction model for sugarcane at field level by integrating time-series remote sensing imagery data with ML techniques. Ground truth agro data and thirty-one spectral vegetation indices were extracted from Sentinel-2 imagery and were considered for yield modeling. A two-level feature selection technique was used to determine the most significant variables that best correlated with sugarcane yield to predict yield in advance. Seven ML algorithms, including those based on regularization, decision trees, and ensemble methods like boosting, were used to predict yield. The approach achieved the highest R2 score of 0.73 and the lowest root mean squared error (RMSE) of 13.45 t/ha with random forest (RF) among the seven ML models tested. Furthermore, all feature selection procedures identified normalized difference red edge (NDRE), red edge chlorophyll index (RECI), and ratio vegetation index (RVI) as major yield-driving variables. The experiments during feature selection demonstrated the potential of red edge spectral bands in development of a reliable sugarcane-yield prediction approach. The RF model obtained using the proposed methodology outperforms the two baseline models developed using NDVI and GNDVI indices, with an improved RMSE of 16-18%.