A spatio-temporal Bayesian Network model: a case study in brazilian Amazon deforestation prediction

• Ano de defesa: 2020
• Autor(a) principal: Alexsandro Cândido de Oliveira Silva
• Orientador(a): Leila Maria Garcia Fonseca
• Banca de defesa: Sidnei João Siqueira Sant'Anna, Solon Venâncio de Carvalho, Carlos Renato Lisboa Francês, Raul Queiroz Feitosa
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Instituto Nacional de Pesquisas Espaciais (INPE)
• Programa de Pós-Graduação: Programa de Pós-Graduação do INPE em Computação Aplicada
• País: BR
• Link de acesso: http://urlib.net/sid.inpe.br/mtc-m21c/2020/05.27.18.15

The key tool for dealing with probabilities in AI is the Bayesian Network (BN). A BN provides a coherent framework for representing and reasoning under uncertainties, which are estimated based on probability theory. However, BNs present some limitations as they do not explicitly model spatial and temporal relationships between variables. Some extensions of BNs have been used to overcome those BNs weaknesses, such as the Spatial BN that integrates GIS and BN and confers to the BN a spatially explicitly strategy, and the Dynamic BN that extends the concept of BNs by relating variables across time. BN approaches have already been proposed to predict LULCC such as deforestation processes. However, deforestation has been considered as a static process when modeled by BNs. In this context, the main goal of this work is to build Spatio-Temporal BN (STBN) models to incorporate both spatial and temporal information in the deforestation risk prediction. For this, we also implemented a package for the R programming language, which enables the development of STBN-based LULCC models for other earth observation applications besides the deforestation process. The STBN models proposed in this thesis are used as a LULCC model for predicting deforestation risk in three priority areas of the Brazilian Legal Amazon: (i) in the southwest of Amazonas State; (ii) in the northwesters of Mato Grosso State; and (iii) surrounding the BR-163 highway in the southwest of Pará State. Among the variables selected to compose the STBN models, the distance from hotspots fires variable stood out as one of the most important for deforestation risk prediction, while protected areas variable was important as a deforestation risk mitigator. The proposed STBN models presented a strong performance with a great agreement between deforestation events and predictions over the years. STBN models results also showed that there was an increase in uncertainty in predictions over time, indicating that more long-term the prediction is, the less accurate it will be. With this, we can state that STBN-based LULCC models are recommended for short-term prediction of deforestation risk.