Registro geométrico e radiométrico automático de imagens Landsat-TM na Amazonia pela maximização da informação mútua

Abstract

A method is described for automatic registration of orbital sensor images obtained at different dates, consisting of both geometric correction and radiometric normalization relative to a reference image. Mutual information is used as the metric of geometric and radiometricsimilarity. The affine transformation is used in the geometric correction and a linear transformation for the radiometric normalization. Powell's method is applied in iterative optimization for both types of registration. The method is validated using images obtained at five different dates by Landsat 's Thematic Mapper (TM) sensor, bands 3, 4 and 5 of scene 231-062 in the Central Amazon. Applied to a 1000x1000 pixels resampled version of the full scene, routines are executed under Linux on a PC with Intel 800 MHz processor. Geometric RMS error ranges 0.9-1.1 pixels for eight control points on the four corrected dates, while radiometric RMS error for populations of pixels whose reflectance is considered temporally invariant ranges 2.2-5.5 DN (TM band 3), 2.4-3.1 DN (band 4) and 2.8-3.6 DN (band 5). The technique is also applied to a time series of images of the area around the Ducke Forest Reserve near the city of Manaus.