Mineração de dados, exatidão da classificação e modelagem do sombreamento do relevo no mapeamento do uso e cobertura da terra

• Autor(a) principal: França, Michelle Milanez
• Data de Publicação: 2013
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: LOCUS Repositório Institucional da UFV
• Texto Completo: http://locus.ufv.br/handle/123456789/1664
• topic: Solos - Classificação; Sensoriamento remoto; Processamento de imagens; Mineração de dados (Computação); Land- Classification; Remote sensing; Image processing; Data mining (Computing); CNPQ::CIENCIAS AGRARIAS::AGRONOMIA::CIENCIA DO SOLO

This work is segmented into three chapters. The first one deals with the quantification of the shadow areas in satellite images, established from existing data about the days and hours of satellite overpass, as well as the establishment of the best times for image acquisition in order to obtain the lowest percentage shadow. The end product of Remote Sensing featuring pixels with considerable areas of shade, especially in the high latitudes of the Northern and Southern Hemispheres. This effect highlights the landforms of relief, but in return affect the work of image classification, preventing it from being obtained class located below the shade. As yet it has not been possible to identify any data to present the area lost by shading in satellite images, it was decided to model the direct solar radiation that reaches the ground, date and time of passage of the Landsat TM and ETM +, whose images are used worldwide. For this, the same conditions were simulated relief at different latitudes, starting from 0° latitude (Ecuador) to 40° S latitude It was found that , in latitude 30° S and 40° S, where the area loss by shading goes from 27 % to 91 % , the images must be acquired, preferably between October and March. At latitudes 0° and 10 ° the loss can be considered negligible when a minimum threshold occurring in 10%. In the second chapter we used two Landsat TM images, acquired in one period of spring/winter, with extensive areas of shadows and over the period of the winter/fall, highlighting a bit of ground targets. These images were submitted to the process of data mining to obtain the best combination of bands and the maximum likelihood algorithm. Data mining was used to check whether there are combinations of bands that add better result in the classification, in addition to traditional 2,3,4 and 345, very widespread in Remote Sensing. In addition to the spectral bands of said sensor, three major components were used and the index NDVI total of 10 bands, combined together, resulting in 1,023 rankings for each image. The highest kappa obtained for the image of spring/winter was 0.90 with combination 235610 and for the winter/fall, up obteu kappa of 0.88 in combination 2358910. Were obtained conditional kappas lower classes only coffee and eucalyptus. When evaluating the effect of the set of bands was found that the main components and vegetation index NDVI increase in kappa provided. The NDVI appeared in all combinations with high kappas. The classifications that had at least one visible band, one of the infrared, one main component and NDVI, showed good results. In the last chapter, it was discussed the question of the accuracy of the maps produced, the indices used to assess the accuracy and limitations, as well as what has been proposed lately. For this, we used the results of the fifty best scores obtained with the two images that were evaluated, pixel by pixel. It was verified that the classes of coffee and eucalyptus were classified into seven to eight categories of diversity. While so much of the catchment area is classified in agreement, ie, these disagreements represent a smaller percentage. Nevertheless, the results showed that the statistical test employed z test, although widely used and recommended to assess statistical differences between binders or combinations of bands was not satisfactory for presenting different results, where, theoretically, there was no statistical difference in the z 5% level of probability. The results of the data showed no difference between responses showed accuracy of 41.4 % for the image of spring / winter and 60.5% for the image of the winter/fall. That is, based on the result obtained only area percentages of these possiam the same response in the classification process.