A Time Series Mining Approach for Agricultural Area Detection

Abstract

Acquiring meaningful data to be employed in building training sets for classification models is a costly task, both in terms of difficult to find suitable samples as well as their quantity. In this sense, Active Learning (AL) improves the training set building by providing an efficient way to select only essential data to be attached to the training set, consequently reducing its size and even enhancing model's accuracy, when compared to random sample selection. In this paper, we proposed a framework for time series classification in order to monitor sugarcane area in São Paulo, Brazil. The AL approach consisted of selecting seasonal time series information from less than 1 percent of each class' pixels to build the training set and evaluate this selection by an expert user supported by distance measurements, repeating this process until both distance measurement thresholds were satisfied. In most years, the classification results presented about 90 percent of correlation with official estimates based on both traditional and satellite image analysis methods. This framework can then help Land Use Change (LUC) monitoring as it produced similar results compared to other methods that demands more human and financial resources to be adopted.