Neste estudo, é investigada a utilização de uma medida chamada dimensão fractal para fins de classificação de imagens digitais. A dimensão fractal é capaz de quantificar as características espaciais das superfícies geradas a partir das imagens de sensoriamento remoto, especialmente a textura destas superfícies. Os valores de dimensão fractal são calculados pixel a pixel, segundo o método dos Prismas Triangu-lares e posteriormente organizados num formato matricial ou raster, em uma estrutura similar a uma ima-gem digital, podendo ser denominados de bandas fractais. Estas bandas fractais podem ser utilizadas de forma semelhante às tradicionais bandas espectrais em classificadores convencionais. Esta hipótese foi testada com uma aerofoto digitalizada da cidade de Porto Alegre - RS. Observou-se que as imagens fractais, embora apresentem qualidade visual inferior, proporcionam maior separabilidade entre as classes presentes e possibilitam a obtenção de índices de acerto maiores nas classificações, quando comparadas com a imagem espectral. USE OF FRACTAL DIMENSION TO SEPARATE TEXTURAL CLASSES IN AN AERIAL PHOTOGRAPH OF THE CITY OF PORTO ALEGRE Abstract Interest is currently growing in the use of spatial attributes for automatic classification of digital images, as is clearly demonstrated by the increasing number of scientific papers on the topic. The reason for this interest is that some classes in natural scenes are not easily distinguished by the spectral features (urban areas, for instance). Urban areas, in particular, are better defined by spatial attributes, such as texture. This research explores the use of fractal dimension to characterize and separate textural classes present in an aerial photograph of Porto Alegre, capital city of the State of Rio Grande do Sul, Brazil. The fractal dimension can be considered as a measure of the spatial complexity of surfaces generated from remotely-sensed images and it is calculated here over moving windows with 7x7 and 9x9 pixels, using the Triangular Prism method. By using a moving window, it was possible to organise the data in a format similar to that used in spectral bands, thus obtaining fractal-dimension bands, which were converted to digital counter values (between 0 and 255). The Bhattacharya distance was used to estimate the separability between pairs of classes, and Gaussian maximum likelihood was used to classify pixels in the images composed of both fractal and spectral bands. The stronger differentiation between classes, together with the high percentage of successes in test samples, shows that the fractal approach can be useful in automatic classification procedures and in situations where the spectral information alone is not sufficient to distinguish the classes successfully.