Change detection in the Brazilian Savanna biome

• Ano de defesa: 2018
• Autor(a) principal: Bueno, Inacio Thomaz
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal de Lavras; Programa de Pós-Graduação em Engenharia Florestal; UFLA; brasil; Departamento de Ciências Florestais
• Palavras-chave em Português: Mudanças na cobertura do solo; Florestas - Sensoriamento remoto; Cerrado - Análise exploratória; Aprendizado de máquina; Land cover changes; Forests - Remote sensing; Cerrado - Exploratory analysis; Machine learning; Recursos Florestais e Engenharia Florestal
• Link de acesso: http://repositorio.ufla.br/jspui/handle/1/29098

Many remote sensing techniques have been developed for forest change detection but there is no optimal method without limitations that can be applied in all landscapes. In the Brazilian savanna biome is not different, the analysis and quantification of human induced deforestation in Cerrado areas proved to be a challenge regarding to the spectral information. This study was divided in two parts, the first one exploring the spectral and temporal information of land cover changes, and in the second we used meaningful information of these changes to discriminate human induced from seasonal changes by different machine learning algorithms. Chapter one evaluated the image data availability in the SF9 basin sampled areas based on cloud and shadows cover, and used filter-based feature selection methods and object-based image analysis to also evaluate Landsat 8 bands. These feature selection methods took red and short wave infrared bands as promisor bands to detect deforestation in savanna biome. In temporal context, free cloud cover presented good change detection accuracies even for distinct image frequencies. Chapter two used the promisor bands previous evaluated to compute spectral indices, which create an input dataset to three machine learning algorithms, Artificial Neural Network (ANN), Random Forest (RF) and Support Vector Machine (SVM), and also assessed individually spectral channels indices in all detections. Random Forest demonstrated the best results in test phase with overall accuracy of 92%. The short wave infrared spectral channel as well as the tasseled cap brightness and greenness transformations indices had positive influence in all machine learning algorithms. Thus, this study emerged new options to savanna change detection through a database exploratory analysis and different machine learning algorithms.