AN EVENT-BASED APPROACH FOR TRAINING DATA SELECTION FOR MAPPING DEFORESTATION

Constructing a good set of training samples is a challenge in machine learning-based classification. This article presents a methodology for selecting samples for mapping deforestation using image time series. We distinguish between event-related samples with time series breaks, seasonal areas, and naturally stable locations. Auxiliary self-organising maps and active learning methods were used to improve sample quality. The data set was used to classify deforested areas in the state of Rondônia in 2022 with an overall accuracy of 0.94 and producer and user accuracies above 90%.