Detecção de mudanças em áreas de cerrado usando inteligência artificial

• Autor(a) principal: Pereira, Eveline Aparecida
• Data de Publicação: 2020
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFLA
• Texto Completo: http://repositorio.ufla.br/jspui/handle/1/45434
• topic: Aprendizado de máquinas; Sensoriamento remoto; Geographic Object-based Image Analysis (GEOBIA); Machine learning; Remote sensing; Cerrado; Recursos Florestais e Engenharia Florestal

Brazil contains large tracts of native vegetation, including large areas of tropical Brazilian Savannas biome, which has been threatened due to the expansion of anthropic activities. In the last years, Remote Sensing (RS) data combined with Artificial Intelligence (AI) have been used to identify the dynamic of the Land use/Land Cover Change (LULCC) of these areas, producing LULCC maps with high accuracy. However, the choice of the AI algorithm and the selection data attributes for the learning process are crucial steps, especially in environments influenced by seasonal variations. Considering these circumstances, the study focus in the following questions: a) what type of attribute (spatial or spectral) or their combination could better differentiate the seasonal changes produced by weather conditions, from atrophic changes in RS images; b) what is the effect of the training sample size into different AI classifiers to produce change maps. Thus, spatial and spectral information were extract for objects generated from Landsat NDVI images in a Tropical Savanna area, acquired at different seasonal periods. The Multilayer Perceptron (MLP), Support Vector Machine (SVM) and Random Forest (RF) algorithms were compared. The MLP produced the most accurate change map, with 75,16% of global accuracy and greater robustness in relation to the variation of the sample intensity. In order to evaluate the generalization capacity of the algorithm, the trained MLP was used to detect changes in contiguous Landsat tiles. The results showed a decrease to 56% of global accuracy, which indicates a limitation of the method. Therefore, the spatial attributes were capable of accurately differentiate deforestation and fires sites, from seasonal changes.