Um método para classificação de imagens de satélite usando Transformada Cosseno Discreta com detecção e remoção de nuvens e sombras

Abstract

This work proposes a supervised algorithm for classi cation of remote sensing images. It is composed by three stages: removal or smoothing of clouds, segmentation and classi cation. The removing clouds method uses homomorphic ltering to deal with obstructions caused by the presence of clouds and the Inpainting method to remove or soften the presence of dense clouds and shadows. The proposed segmentation and classi cation approaches are based on AC power coe cients of the Discrete Cosine Transform (DCT). Classi cation is used in the supervised mode. An image database is used to evaluate the implemented algorithm. This database is composed by 14 images obtained from various sensors, which 12 have some kind of obstruction. The Peak Signal-to-Noise Ratio (PSNR) and Kappa coe cient metrics are used to evaluate the removal or smoothing of clouds and shadows method. In this stage, several high-pass lters were compared to choose the most e cient. The image segmentation task is evaluated by the Edge Border Con dence (EBC) and the classi cation task is evaluated by the measure of the relative entropy and by the mean squared error (MSE). The resulting images are presented to allow the subjective evaluation by visual comparison. The experimental results show the e ciency of the proposed algorithm, especially when compared to the Spring software, distributed by the Instituto Nacional de Pesquisas Espaciais (INPE).