Classification of sugarcane varieties with harmonized Sentinel-2 and Landsat-8/9 data using parametric and non-parametric methods

Abstract

Remote sensing data has been successfully used to enhance sugarcane monitoring and management, in topics such as yield estimation, health anomaly detection, or variety classification. Specifically, variety classification is an essential objective for optimizing crop management, as it can guide strategies such as plant renovation, pest control, or yield estimation. A literature review allowed identifying that the integration of diverse satellite platforms to enhance time series for sugarcane variety classification has not been explored. This strategy can improve the temporal density of available imagery in our study area, Costa Rica, with frequent cloud cover. Therefore, our research proposed to classify six sugarcane varieties using an additive approach (aggregating them in four variety groups) and employing parametric and non-parametric algorithms on harmonized data from Sentinel-2 and Landsat-8/9. Validation was done at both pixel and plot scales. The best classifications were achieved using green and near infrared bands, along with the Enhanced Bloom Index and Normalized Difference Infrared Index vegetation indices. Regarding temporal dynamics, the most relevant months were September, November, and December, corresponding to advanced growth cycle stages. Support Vector Machine and Random Forest provided the best classification accuracies. At the pixel scale, the overall accuracy of all groups exceeded 0.86, with a slight decrease as the number of varieties increased. When validation was done at plot scale, the overall accuracy remained above 0.89 in all the groups. These achievements were suitable and valuable for sugarcane sustainable planning and decision-making.