Machine Learning Models to Estimate the Sugarcane Brix Values from Multitemporal Vegetation Indices

Remotely sensed data in combination with image processing and geoscience is essential to support decision making related to the crop productivity and assessment of the crop quality in terms of its growth. Continuous improvements in accuracy by eliminating all the radiometric and geometric errors, as well as the use of sophisticated statistical techniques goad to effective and optimized decision making. Although previous studies found mathematical relationships between final crop yield and different forms of spectral indices all around the world, few studies conducted in small areas. The relation of crop growth stages and sugarcane yield after applying the atmospheric and geometric corrections is also not considered by the researchers in the past. This study aims to estimate the sugar content of the sugarcane crop grown in a small area with machine learning models. The process based on correlation analysis to identify of the best periods for the reliable and efficient estimation of the crops is presented in this study. The study has been carried out in the Khedli Minor area near Bahadrabad District Haridwar Uttarakhand, which belongs to the Himalayan foothill region. The performance evaluation of the models has been analyzed using Mean Absolute Error and Root Mean Square Error (RMSE). The minimum value of RMSE =1.52 has been found with the cost parameter C=1 and Sigma = 0.1 for the Support Vector Regression (SVR) model. Compared with the other models, the SVR model exhibit better performances on Brix estimation based on the multitemporal vegetation indices. Although, this study exhibits some promising results in the yield estimation, a more generalized model can be developed with additional inputs from physical characteristics such as temperature, precipitation and soil moisture.