Detecção de diferentes alvos no entorno de reservatórios no semiárido através do uso de sensoriamento remoto

Detalhes bibliográficos
Ano de defesa: 2017
Autor(a) principal: Araújo, Efraim Martins
Orientador(a): Não Informado pela instituição
Banca de defesa: Não Informado pela instituição
Tipo de documento: Tese
Tipo de acesso: Acesso aberto
Idioma: por
Instituição de defesa: Não Informado pela instituição
Programa de Pós-Graduação: Não Informado pela instituição
Departamento: Não Informado pela instituição
País: Não Informado pela instituição
Palavras-chave em Português:
Link de acesso: http://www.repositorio.ufc.br/handle/riufc/22929
Resumo: The main goal of this work is to evaluate the potential of discrimination for soil use and occupation in the surroundings of reservoirs located in the semi-arid region, using spectral information obtained by remote sensor considering multispectral and hyperspectral satellites images. The satellite images selected for the survey are Landsat 8 and Hyperion images. The research evaluated and compared the performance of different techniques for image classification applied to multispectral (Landsat 8) and hyperspectral (Hyperion) sensors aiming the detection and delineation of the land uses around the reservoirs Paus Brancos, Nova Vida and Marengo, located in the 25 de Maio settlement, Madalena – CE, belongin the hydrographic basin of the Banabuiú reservoir. The classes identified based on surveys conducted in 2014 and 2015 campaigns around the reservoirs were: water (water bodies), macrophytes, exposed soil, native vegetation, agriculture, sparse vegetation and fload plaind crop, in addition to cloud and shadow targets. Different techniques for image processing are tested and compared, such as NDVI (Vegetation Index by Normative Difference), non-supervised classifier (ISODATA) and supervised classifiers (Maximum Likelihood, K-Nearest Neighbours - KNN, Minimal Distance and Random Forest). For processing hyperspectral images, we use SVM (Support Vector Machine) classifier, which provides to analyze all the 155 radiometrically calibrated bands of the Hyperion sensor, assigning them weights in the classification process. According to the results provided by SVM classifier, RGB compositions of the 10 best ranked bands are evaluated aiming the identification of the best successful combination for delineating classes in the surroundings of the three studied reservoirs (bands R – 51, G – 161, B – 19). The analysis of NDVI multispectral images behaved inaccurate for delineating classes, mainly considering targets with similar spectral response, such as some kinds of vegetation. Meanwhile, the unsupervised classification proved to be deficient, not being able to discriminate water bodies from cloud shadow, even after applying contrast enhancing techniques within the Matlab computing program environment. The spectral and temporal analysis of soil use reflectance allowed to identify the spectral behavior of the nine classes considered in this study and also the spectral bands with the highest potential for discriminating the referred classes. Indeed, even within these optimal bands, some targets present similar spectral behaviors, difficulting their discrimination. On the other hand, the supervised classification applied to Landsat 8 and Hyperion images achieved to be succeed in the delineation of either distinct (water, soil and vegetation) and similar (macrophytes, fload plaind crop, native vegetation, agriculture and sparse vegetation) targets. It should be emphasized that the performance results of the classifiers applied to the Hyperion images are generally superior to those obtained respectively by the same classifiers over the Landsat 8 images. This can be explained by the higher spectral resolution of the first sensor, which increases the potential for delineating targets with similar spectral response. Concerning the supervised classifiers, in the stage of performance test, it was observed that KNN method is more accurate than the others for Landsat 8 images, with a maximum Kappa coefficient equal to 0.68. Meanwhile, for Hyperion images, the Maximum Likelihood method achieves the highest performance result, with a maximum Kappa coefficient equal to 0,78. Additionally, a sensitivity analysis of the supervised classification applied to Landsat 8 and Hyperion images is performed regarding the number of samples per class randomly collected for training. It is clearly observed that the randomness concerning training stage allows finding subsets of samples which increase the performance results. For the evaluation of the supervised maximum likelihood classification method, Landsat 8 (24/08/2015) and Hyperion (285/08/2015) images are considered for the computing tests. The training data were collected through a research technical visit in November, 2015, around São Nicolau reservoir, also located in the 25 de Maio settlement, while the data for performance evaluation (validation) were extracted from the image generated through the overflight performed by an Unmanned Aerial Vehicle (UAV), in the same period in the Paus Brancos reservoir. The obtained results demonstrate the robustness for that classifier when applied to Hyperion image, with a Kappa of 0.83. Concerning Landsat 8 image, the computed Kappa is 0.49, which can be explained by the corresponding lower spectral resolution. Two other applications of the Maximum Likelihood classifier for Landsat 8 and Hyperion images were performed. In the first one, the accuracy of each classifier for detecting reservoirs contours was tested. In some of these reservoirs, that task is made difficult by the presence of macrophytes in the hydraulic basin. For this analysis, the intersection area between the scenes of the Landsat 8 and Hyperion sensors, which cover the area of 25 de Maio Settlement, was used, totalizing 48 reservoirs. The results showed that the classifier generally underestimates the reservoir areas, reaching 73% and 51% of the reference value in the Landsat 8 and Hyperion images, respectively. Finally, an application of the supervised Maximum Likelihood classifier was performed using Hyperion images for the detection of land uses in the surroundings of reservoirs of other regions of the State of Ceará. In the analysis of the available data, it is possible to identify a reservoir located in the municipality of Lavras da Mangabeira, displayed in the Hyperion image (26/09/2010), with low cloud cover, near the image of Google Earth (08/07/2009), also used for validation purposes. The results of the application indicate accurate performance for the classifier associated with the RGB composition selected for the Hyperion image (bands R - 51, G - 161, B - 19) concerning the detection of the uses around this reservoir, the resultant Kappa coefficient is 0.90. On the other hand, the availability of Hyperion sensor data in applications for the State of Ceará is very restricted, which makes difficult to develop continuous researches using hyperspectral images.