Avaliação de descritores texturais geoestatísticos e de Haralick para o reconhecimento de plantas daninhas

• Autor(a) principal: Barbosa, Danilo Pereira
• Data de Publicação: 2009
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: LOCUS Repositório Institucional da UFV
• Texto Completo: http://locus.ufv.br/handle/123456789/4023
• topic: Reconhecimento de padrão; Geoestatística; Análise discriminante; Textura; Recognition pattern; Geoestatistics; Discriminant analysis; Texture; CNPQ::CIENCIAS AGRARIAS

The concern in minimizing the amount of chemical products used in farmings is increasing. The use of artificial vision systems has been demonstrating a great potential for use of varied taxes of inputs, as for instance, the application herbicides only in places where the presence of harmful plant is detected. The good acting of a system developed for this purpose depends mainly of the characteristic use that they allow to differentiate patterns of harmful plants of the pattern of the cultivated species. Like this, the objective of the present work was to develop and to evaluate a characteristic for the recognition of the patterns corn plant and harmful plant. The specific objectives were: the) to identify which image, green excess or the index of vegetation of normalized green, tends to provide better classification; b) to compare the classification obtained by characteristics geoestatistics, obtained when using the characteristics of Haralick. With this purpose, were acquired to the 29 days after the emergency, period in that it is usually made the application of herbicides, nine corn images (Zea Mays L.) and of each one of the species of appraised harmful plants in this experiment: Euphorbia heterophylla L., Digitaria horizontalis Willd, Cenchrus chinatus L. Six of these images were used for the selection of the characteristic that promotes better acting in the classification. The remaining three were used for the validation of the selected characteristic. Each one of the six training images was cut out in 100 blocks of 68x68 pixels. For each one of the blocks was obtained the value of the characteristic textural geoestatistic (variogram, the madogram, cross variogram and pseudo cross variogram) and the one of Haralick (angular moment, average, variance, entropy, correlation, moment of the product, inverse moment of the difference and correlation measures). Additionally, characteristic geoestatísticos and no-geoestatísticos they were obtained considering different angles (0, 45, 90 and 135°) of relationship among pixels. Characteristic geoestatistics were, also, obtained for different distances (1, 2, 3, 4, 5, 6, 7, 8, 9, 10) of in pairs among pixels. The characteristic variograma and madograma were calculated to leave of the image green excess and GNDVI. Already the characteristic cross variogram and pseudo cross variogram were calculated with Greenness Method use in the blocks using the combinations of the bands RxG, GxB and IVxG. The characteristic of Haralick were calculated starting from the images of the green excess and GNDVI. The acting of the characteristic, proposed like this, it was evaluated using discriminate analysis. The selected characteristic were those that presented larger value for the index kappa. Additionally, new characteristic were obtained starting from combinations of the selected characteristic. These combinations, also, had appraised acting using the discriminant analysis with the objective of to identify which combination provides better classification. Later, the power of generalization of the selected combination was evaluated using the three images of each species reserved for the validation stage. The conclusions obtained regarding the objectives proposed in this research were a) the image that tended to present the best results of the index kappa was the image excess of green; b) the characteristic obtained starting from the function madograma and the one of Haralick were the ones that supplied the best results; c) the characteristic geoestatistic madograma in the 10 distances and angle 0° presented better classification results when used without combination of other characteristic; d) the characteristic geoestatísticos and the one of Haralick, when used separately didn't present such good as combined results; e) the characteristic use that consider the continuity of the pixel values, in the recognition of patterns can be a fundamental tool in the classification process.