LANDUSE/LANDCOVER CLASSIFICATION BY NON-PARAMETRIC ALGORITHMS COMPARED TO THE LIKELIHOOD CLASSIFIER

Abstract

Landuse/landcover maps produced by non-parametric classifiers of a Landsat TM image were compared with results obtained by the usual likelihood classifier (MAXLIKE). The maximum likelihood classifier is the parametric strategy pixel the pixel most used in orbital images classification. But Skidmore and Turner (1988) obtained results of global accuracy with their algorithm non parametric in SPOT XS data, for classes of Pinus spp. age superior 14% in relation to the algorithm of maximum likelihood. The algorithm of Skidmore/Turner suffered modification proposal in Lowell (1989), Gong and Dunlop (1991) and Dymond (1993). The purpose of the algorithm non-parametric that were developed in this work was to diminish the limitation of the Skidmore/Turner non-parametric classifier that, compared to MAXLIKE, demands a very large training sample to reduce the not classified areas in the image. This algorithm, a new supervised nonparametric classifier that make flexible the association amidst brightness values was developed for testing the assumption that it increases the image classified area without missing accuracy and for solving the limitation of the nonparametric classifier of Skidmore/Turner, that requires a larger training sample for reducing the of unclassified pixels in the image. To facilitate the understanding of the algorithms used in this work, a simplified example of data of remote sensing is presented, which corresponds to a window of image of 20 x 20 pixels, with two classes sampled: lagoon and forest, according to the training samples in Table 1, extracted from 3,4 and 5 Landsat TM bands. The main difference of the nonparametric classifier of Skidmore/Turner in relation to MAXLIKE is in p(X|j) calculation. The training samples to that classifier are constituted of discrete groups of pixels vectors, common to each class, and no more a continuous space determined by the multivariate normal function. For the purpose of reducing the restrictive effect of non-parametric of Skidmore/ Turner, NPVIC it was developed a supervised non-parametric algorithm, by pixel, which had on principle to make flexible the rigid association among brightness values of the pixel vector. This algorithm developed from the conditional probability P(i|X), considers the brightness values independent for band and attributes to the pixel the class that integrates the maximum value of intersections by bands, pondered by the sizes of the classes. The area is placed in the microregion of Viçosa, Minas Gerais, Brazil, with predominance in this municipal district. Part of an image was used of 15 x 15 km (22.500 ha), extracted from the Landsat TM image, 217/74, South Quadrant, of 10/10/94, composed of TM 3, 4 and 5 band. Eleven (11) classes of landuse and landcover were defined, described in Table 8. The reference sample was constituted by a systematic grid of 163 points, distant of 28 pixels on the area of the photos, more 146 intermediary points among the points of the bars (total of 309 points). The index used to evaluate the results of the classifications was the global accuracy (G). The non-parametric algorithm of Skidmore/Turner, the non-parametric algorithm of Skidmore/Turner modified by Dymond, NPVIC, and NPVIC with of Dymonds normalization (Apendix) were written in-house software in Pascal Language, utilizing modules of selection of training samples of the IDRISI 2.0 for Windows (vectorization Screen and MAKESIG). The algorithm of maximum likelihood used was MAXLIKE of the same software. The results of global accuracy and the % of the classified area in the image are presented in Table 9 and the Figure 4 shows the ranking of results of G and %ANC to all the classifiers and strategies applied to the compression factors and original data, as well as the test Z of seqüencial significance among the results 1%. The combination of the best accuracy results with the smallest percentages of non-classified area occurs to the algorithm of maximum likelihood with options 0 and 1%. The largest levels of %ANC are represented by the algorithms NPSKID, NPSKIDYM and NPVICB3. And the smallest indexes of accuracy (G) are represented by the algorithms NPVICA1 and NPVICDYMA1. The non-parametric NPVIC algorithm allows to reduce the non-classified area in an image, when compared to NPSKID, but not in a satisfactory way, since for classifying the whole area images of low accuracy. The Skidmore/Turners non-parametric algorithm to sizes of training sample sufficient to MAXLIKE, produces classifications of low quality. The results showed the superiority of MAXLIKE regarding accuracy and not classified area in the image. In any way, non-parametric classification does not need the presupposition of normality of data hence it uses real information from training samples for classification that ensures bigger accuracy.