Classification Analysis of NDVI Time Series in Metric Spaces for Sugarcane Identification

In Brazil, agribusiness is an important task for the economy, since it provides a substantial part of the country’s Gross Domestic Product. Besides that, interest in biofuels has grown, considering they make viable the use of renewable energy. Brazil is the world’s largest producer of sugarcane, which enables a large ethanol production. Thus, to monitor agricultural areas is important to support decision making. However, the amount of generated and stored data about these areas has been increasing in such a way that far exceeds the human capacity to manually analyze and extract information from it. That is why automatic and scalable data mining approaches are necessary. This work focuses on the sugarcane classification task, taking as input NDVI time series extracted from remote sensing images. Existing related works propose to analyze non-metric features spaces using the DTW distance function as a basis. Here we demonstrate that analyzing the multidimensional space with Minkowski distance provides better results, considering a variety of classifiers. kNN using L2 distance performed similarly or better than using DTW. We also demonstrate a data configuration with geolocation for training XGBoost, with results better than state-of-the-art.