INTEGRAÇÃO DA INCERTEZA NA AMOSTRAGEM E CLASSIFICAÇÃO RANDOM FOREST UTILIZANDO BANDAS E ÍNDICES ESPECTRAIS PARA O MAPEAMENTO DE INUNDAÇÃO
Integration of uncertainty in sampling and random forest classification using bands and spectral indices for flood mapping
DOI:
https://doi.org/10.5016/geociencias.v41i04.16802Abstract
Traditional classifications present limitations for mapping floods due to mixing the spectral response of water with adjacent non-aquatic targets or similar spectral response of non-aquatic targets with water. Furthermore, in general, these classifications are evaluated only in terms of overall accuracy without considering the uncertainties in the classification process. Thus, this study aimed to integrate uncertainty in the Random Forest (RF) classification process for flood mapping, which guided the sampling process. The classification used 21 variables including indices and spectral bands from the Operational Land Imager sensor of the Landsat-8 satellite. Sampling was performed initially with the selection of points from the visual interpretation of the satellite image and later by collecting samples with high Shannon entropy values in the uncertainty map. The variables with the greatest importance for classification were selected by the Recursive Feature Elimination (RFE) algorithm. The final RF classification using samples collected based on the uncertainty map and with the four selected variables by the RFE presented an accuracy of 98.0% and a reduction of uncertainty, which indicates a greater confidence in the spatial representation and quantification of water permanent and temporary surface associated with floods.
Keywords: Flood mapping. Random Forest Classifier. Spectral bands and indices. Variable selection. Shannon Entropy.
