Sugarcane Yield Prediction Using Vegetation Indices in Northern Karnataka, India

Abstract

The integration of remote sensing (RS) technology with machine learning (ML) algorithms can facilitate accurate prediction of sugarcane yield. This paper presents an assessment of the random forest (RF)-based prediction model and second-degree polynomial regression models for sugarcane yield prediction. The models are developed utilizing vegetation indices (VIs) computed from the Sentinel-2 satellite and sugarcane yield data. The sugarcane yield data were acquired from sugarcane fields around the Godavari Bio-refineries Limited (GBL) factory in Karnataka, India, during the 2017–2018 sugarcane growing season. A dataset detailing agronomic information and VIs was prepared for yield prediction. The study area comprises seven sugarcane growing talukas. The second-degree polynomial regression was used for predicting the sugarcane yield as it had the best fit for the distribution of variables. The green normalized difference vegetation index (GNDVI) recorded the highest R2, i.e., 0.71 during November month with a coefficient of variance of 0.83, with all other indices characterized by R2 values ranging from 0.42(modified chlorophyll absorption ratio index) to 0.69 (normalized difference red edge), suggesting the GNDVI’s potential for sugarcane yield prediction. Comparing the actual yield with the predicted yield, the RF prediction and second-degree polynomial regression model exhibited accuracies of 90.42% and 88%, respectively. This indicates that the models are sufficiently accurate and beneficial in decision-making for sugar mill operational planning.