EXPLORING THE USE OF CLASSIFICATION UNCERTAINTY TO IMPROVE CLASSIFICATION ACCURACY

Abstract. Supervised classification of remotely sensed images has been widely used to map land cover and use. Since the performance supervised methods depends on quality training data, it is essential develop generate an enhanced dataset. Active learning represents alternative for such purpose as proposes create a dataset optimized samples, normally collected based uncertainty. However, heavily dependent human interaction, since user label selected samples over number iterations. In this paper, we explore use uncertainty improve accuracy through single iteration. We conducted experiments in region Portugal (Trás-os-Montes), using multi-temporal Sentinel-2 images. The proposed approach consisted computing Random Forest collect additional data from areas high perform new classification. An assessment was performed compare overall initial classifications. results exhibited increase accuracy, though considered not statistically significant. Obstacles related labelling sampling units resulted lack various classes, which might have limited improvement. Additionally, uneven proportion per class collection sample regions also prevented higher accuracy. Nevertheless, visual inspection maps revealed that reduced confusion between some classes.