Classificação automatizada do uso e cobertura do solo utilizando imagem Landsat no Município de Araponga, MG
Data
2009
Autores
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Federal de Viçosa
Resumo
Planejar o espaço na busca de entender o presente com informações do passado, bem como projetar cenários futuros é de fundamental importância para a melhor gestão ambiental. Neste sentido, dentro de um programa mais amplo de monitoramento, a detecção da cobertura florestal - de forma ágil, rápida e eficiente - pode contribuir para uma melhoria na proteção e qualidade ambiental, principalmente, em municípios que se destacam por apresentarem biodiversidade relevante, em especial àqueles que são sede de alguma Unidade de Conservação (UC). O uso de imagens de satélite tem se intensificado nos últimos anos, e incrementa o arsenal de informações disponíveis sobre o meio natural. Entretanto, a maior parte da tecnologia disponível para a detecção automatizada, em especial os softwares utilizados, apresentam elevado custo financeiro de aquisição. Neste contexto, o objetivo do presente trabalho foi avaliar novas tecnologias e metodologias de detecção automatizada da cobertura vegetal em imagens orbitais. Para tanto, foram executadas as classificações pelos métodos da Máxima Verossimilhança (MAXVER), das Redes Neurais Artificiais (RNA) e das Árvores de Decisão, utilizando-se de tecnologias de caráter gratuito e, ou, já disponíveis no Instituto Estadual de Florestas (IEF-MG). Essa iniciativa também visa contribuir na avaliação de novas soluções para os problemas e desafios relacionados ao mapeamento de cobertura vegetal enfrentados pelo setor de monitoramento ambiental do IEFMG, em especial pelo Centro de Estudos e Desenvolvimento Florestal, ligado à Gerência de Monitoramento e Geoprocessamento. Para a efetivação do presente projeto foram utilizados as bandas 1, 2, 3, 4, 5 e 7 da cena 217/74 de imagens do sensor LandSat 5 TM (Thematic Mapper), tomadas em 16/11/2005 (verão), correspondente ao município de Araponga (MG), onde concentra-se a maior área do Parque Estadual da Serra do Brigadeiro (PESB). Essa mesma cena foi utilizada para a produção do documento: “Mapeamento da Flora Nativa e Reflorestamento de Minas Gerais” (MFNR-MG), atualmente um referencial para a gestão ambiental do Estado de Minas Gerais. Foram consideradas nesse estudo as seguintes classes de vegetação: Floresta Estacional Semidecídua, Floresta Ombrófila, Campo de Altitude, Pastagem, Café e Eucalipto. Para a classificação automatizada das imagens, inicialmente os dados foram preparados com os Sistemas de Informações Geográficas (SIG): Sistema de Processamento de Informações Georreferenciadas (SPRING) e ArcGis 9.0. Posteriormente, foi realizada a classificação das imagens pelos métodos de MAXVER e o das RNA’s. O SIG SPRING foi usado na classificação por MAXVER e, para as RNA’s, foi utilizado o software Stuttgart Neural Network Simulator (SNNS). Ao final dos trabalhos de classificação, adicionalmente foi realizada uma comparação com os resultados alcançados no MFNR-MG, no qual outro método de classificação, o de Árvores de Decisão, foi empregado em seu procedimento de análise. Os resultados obtidos indicaram que a metodologia por MAXVER, por mais que tenha gerado confusão entre algumas classes por considerar somente o valor da reflectância, o que dificultou ou mesmo impediu a diferenciação entre algumas tipologias vegetacionais, em especial, entre campo de altitude e afloramentos rochosos; e entre eucalipto e floresta semidecídua, apresentou um bom resultado, atingindo um desempenho geral de aproximadamente 80%. A classificação efetuada por RNA’s também não distinguiu efetivamente todas as classes pretendidas, mesmo considerando o plano de informação altitude, segundo fornecido pelo modelo de elevação do terreno. A comparação, dos classificadores testados no presente trabalho com a metodologia adotada no MFNR-MG indicou diferenças expressivas na quantificação das coberturas vegetacionais da área estudada, em especial quanto à formação Florestal Ombrófila, a qual se apresentou bem mais evidente nas classificações aqui executadas (MAXVER e RNA). Embora a metodologia de RNA’s seja amplamente aceita como a mais adequada para a classificação de imagens de satélites, a complexidade e o tempo demandado na preparação dos materiais, bem como os vários procedimentos de tentativa e erro requeridos para sua execução dificultam ou mesmo restringem sua utilização, principalmente na demanda comercial. Por sua vez, dada a simplicidade e os resultados alcançados, a classificação por MAXVER desponta como uma opção mais viável em muitas situações, tais como às classes que não são distinguidas por algum outro fator que não seja a reflectância da imagem utilizada. No entanto, as duas metodologias aqui testadas (RNA e MAXVER), bem como a utilizada no MFNR-MG, não apresentaram bons resultados para as duas classes de cobertura vegetal exótica pretendida neste trabalho (eucalipto e café).
The projection of future scenarios and the space planning, in the search to understand the present, with information from the past, are extremely important for a better environmental management. In this sense, inside a broader monitoring program, the detection of the forest coverage in an agile, fast and efficient way, may contribute to improve the environmental protection and quality, especially in cities that present relevant biodiversity or comprise any Conservation Unit (CU). The use of satellite images has been intensified over the last years, thus providing more information about the natural environment. However, most part of the technology available for automated detection is very expensive, mainly the software systems used. Therefore, the objective of the present work was to evaluate new technologies and methodologies for the automated detection of plant coverage in orbital images. For such, classifications were performed by the methods of Maximum Likelihood (MLE), Artificial Neural Networks (ANN) and Decision Tree Method (DTM), using free technologies and/or technologies already available at the Instituto Estadual de Florestas (Forest State Institute) (IEF - MG). This initiative also intends to contribute for the evaluation of new solutions for the problems and challenges related to the mapping of the plant coverage faced by the environmental monitoring sector of the IEF-MG, mainly by the Centro de Estudos e Desenvolvimento Florestal (Forest Study and Development Center), linked to the Monitoring Management and Geoprocessing. For the execution of the present project, the bands 1, 2, 3, 4, 5 and 7 of the scene 217/74 of images of the LandSat 5 TM (Thematic Mapper) sensor, taken on 11/16/2005 (summer) were used, corresponding to the city of Araponga (MG), where most part of the area of Parque Estadual da Serra do Brigadeiro (Serra do Brigadeiro State Park) (PESB) is concentrated. This same scene was used for the production of a document entitled: “Mapeamento da Flora Nativa e Reflorestamento de Minas Gerais” (Mapping of the Native Flora and Reforestation of Minas Gerais) (MFNR-MG), which is nowadays a reference for the environmental management in the state of Minas Gerais. The following vegetation classes were considered in this study: Semidecidual Stational Forest, Ombrophile Forest, Altitude Field, Pasture, Coffee and Eucalyptus. For the automated classification of the images, the data were initially prepared with the Geographic Information Systems (GIS): Georeferenced Information Processing System (GIPS) and ArcGis 9.0. Later, it was performed the classification of the images by the MLE and ANN methods. The GIS GIPS method was used in the classification by MLE while, for the ANN, it was used the Stuttgart Neural Network Simulator (SNNS) software system. In the end of the classification works, it was carried out an additional comparison with the results achieved in the MFNR-MG, with the use of the Decision Tree Method for the analysis. The results achieved indicated that the MLE methodology presented a good result, with a general performance of 80%, although it generated some disorder between some classes for considering only the reflectance value, thus hindering or even preventing the differentiation between some plant typologies, especially between altitude field and rock outcrops; and between eucalyptus and semidecidual forest. The classification carried out by ANN did not effectively distinguish all the desired classes either, even considering the altitude information plan, as provided by the model of the terrain elevation. The comparison between the classifiers tested in the present work and the methodology adopted in the MFNR-MG indicated great differences in the quantification of the plant coverages of the area studied, especially as to the Ombrophile Forest formation, which was much more evident in the classifications carried out here (MLE and ANN). Although the ANN methodology is widely accepted as the most adequate for the classification of satellite images, its complexity and the time demanded in material preparation, as well as several procedures of trial and error required for its execution, hinder or even prevent its use, mainly in commercial demand. On the other hand, due to its simplicity and the results achieved, the MLE classification turns out to be the most viable option in many situations, such as when classes cannot be distinguished by another factor but the reflectance of the image used. However, both methodologies tested (ANN and MLE), as well as the one used in MFNR-MG, did not present good results for both classes of exotic plant coverage intended in this work (eucalyptus and coffee).
The projection of future scenarios and the space planning, in the search to understand the present, with information from the past, are extremely important for a better environmental management. In this sense, inside a broader monitoring program, the detection of the forest coverage in an agile, fast and efficient way, may contribute to improve the environmental protection and quality, especially in cities that present relevant biodiversity or comprise any Conservation Unit (CU). The use of satellite images has been intensified over the last years, thus providing more information about the natural environment. However, most part of the technology available for automated detection is very expensive, mainly the software systems used. Therefore, the objective of the present work was to evaluate new technologies and methodologies for the automated detection of plant coverage in orbital images. For such, classifications were performed by the methods of Maximum Likelihood (MLE), Artificial Neural Networks (ANN) and Decision Tree Method (DTM), using free technologies and/or technologies already available at the Instituto Estadual de Florestas (Forest State Institute) (IEF - MG). This initiative also intends to contribute for the evaluation of new solutions for the problems and challenges related to the mapping of the plant coverage faced by the environmental monitoring sector of the IEF-MG, mainly by the Centro de Estudos e Desenvolvimento Florestal (Forest Study and Development Center), linked to the Monitoring Management and Geoprocessing. For the execution of the present project, the bands 1, 2, 3, 4, 5 and 7 of the scene 217/74 of images of the LandSat 5 TM (Thematic Mapper) sensor, taken on 11/16/2005 (summer) were used, corresponding to the city of Araponga (MG), where most part of the area of Parque Estadual da Serra do Brigadeiro (Serra do Brigadeiro State Park) (PESB) is concentrated. This same scene was used for the production of a document entitled: “Mapeamento da Flora Nativa e Reflorestamento de Minas Gerais” (Mapping of the Native Flora and Reforestation of Minas Gerais) (MFNR-MG), which is nowadays a reference for the environmental management in the state of Minas Gerais. The following vegetation classes were considered in this study: Semidecidual Stational Forest, Ombrophile Forest, Altitude Field, Pasture, Coffee and Eucalyptus. For the automated classification of the images, the data were initially prepared with the Geographic Information Systems (GIS): Georeferenced Information Processing System (GIPS) and ArcGis 9.0. Later, it was performed the classification of the images by the MLE and ANN methods. The GIS GIPS method was used in the classification by MLE while, for the ANN, it was used the Stuttgart Neural Network Simulator (SNNS) software system. In the end of the classification works, it was carried out an additional comparison with the results achieved in the MFNR-MG, with the use of the Decision Tree Method for the analysis. The results achieved indicated that the MLE methodology presented a good result, with a general performance of 80%, although it generated some disorder between some classes for considering only the reflectance value, thus hindering or even preventing the differentiation between some plant typologies, especially between altitude field and rock outcrops; and between eucalyptus and semidecidual forest. The classification carried out by ANN did not effectively distinguish all the desired classes either, even considering the altitude information plan, as provided by the model of the terrain elevation. The comparison between the classifiers tested in the present work and the methodology adopted in the MFNR-MG indicated great differences in the quantification of the plant coverages of the area studied, especially as to the Ombrophile Forest formation, which was much more evident in the classifications carried out here (MLE and ANN). Although the ANN methodology is widely accepted as the most adequate for the classification of satellite images, its complexity and the time demanded in material preparation, as well as several procedures of trial and error required for its execution, hinder or even prevent its use, mainly in commercial demand. On the other hand, due to its simplicity and the results achieved, the MLE classification turns out to be the most viable option in many situations, such as when classes cannot be distinguished by another factor but the reflectance of the image used. However, both methodologies tested (ANN and MLE), as well as the one used in MFNR-MG, did not present good results for both classes of exotic plant coverage intended in this work (eucalyptus and coffee).
Descrição
Dissertação de Mestrado defendida na Universidade Federal de Viçosa
Palavras-chave
Classificação automatizada; Cobertura vegetal; Imagem de satélite;, Automated classification; Forest coverage; Satellite images;
Citação
Moreira, Gilberto Fialho. Classificação automatizada do uso e cobertura do solo utilizando imagem Landsat no Município de Araponga, MG. Viçosa, MG : UFV, 2009. 85 f. : il. (Dissertação - Mestrado em Solos e Nutrição de Plantas) - Universidade Federal de Viçosa. Orientador: Raphael Bragança Alves Fernandes. T631.4098151 M838c 2009