Publicação
Métodos probabilísticos e estatísticos na gestão de fogos florestais
| Resumo: | Forest fires can be associated to spatial coordinates (longitude and latitude) and to time of occurrence, thus forest fires can be seen as a spatio-temporal point pattern. And if each location has information about the area burned, then we have a marked spatio-temporal point pattern. The aim of this work is to analyze the spatio-temporal patterns of forest fires in Portugal. The data set consists of satellite imagery records of 13457 forest fires larger ar equal to 35 hectares, observed in Portugal through the years 1975 to 2005. Firstly we began by performing a descriptive analysis of the data using the theary of spatial point process and extend to the spatiotemporal point processes and to the marked spatial processes. Then, we modeled the location of forest fires by a log-Gaussian Cox process and we used this model to model the area burned. This way, we tried to show how the point configuration affects the dimension of the fire. To validate the log-Gaussian Cox model we used the notion of spatial residuais and the L-function based on Ripley's K-function. Based on these models, two maps were created: a fire risk map where the risk of fire is associated to the density of fires and a fire danger map where the danger of fire is associated to the area burned. Lastly, the data was transformed into incidence data, binary data indicating the presence of fire, and fire risk maps were built based on transition probabilities (IvIarkov model) and the hazard function of the survival models (discrete Weibull model). In all models were considered several covariates. Regarding the results obtained, the models allow to identify areas where more fires occurred as well as the largest fires, however the implemented models based on space-time point processes underestimates the density and size of fires that actually occurred. The inference on these mo deIs were carried using Bayesian hierarchical models, the SPDE approach and the IN LA approach, except in the discrete Weibull model where the inference was carried via the simulation methods IvICIvIC through WinBUGS. |
|---|---|
| Autores principais: | Pereira, Paula |
| Assunto: | Estatística - probabilidades Incêndios florestais Modelos bayesianos Portugal Teses de doutoramento - 2014 |
| Ano: | 2014 |
| País: | Portugal |
| Tipo de documento: | tese de doutoramento |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório da Universidade de Lisboa |
| Resumo: | Forest fires can be associated to spatial coordinates (longitude and latitude) and to time of occurrence, thus forest fires can be seen as a spatio-temporal point pattern. And if each location has information about the area burned, then we have a marked spatio-temporal point pattern. The aim of this work is to analyze the spatio-temporal patterns of forest fires in Portugal. The data set consists of satellite imagery records of 13457 forest fires larger ar equal to 35 hectares, observed in Portugal through the years 1975 to 2005. Firstly we began by performing a descriptive analysis of the data using the theary of spatial point process and extend to the spatiotemporal point processes and to the marked spatial processes. Then, we modeled the location of forest fires by a log-Gaussian Cox process and we used this model to model the area burned. This way, we tried to show how the point configuration affects the dimension of the fire. To validate the log-Gaussian Cox model we used the notion of spatial residuais and the L-function based on Ripley's K-function. Based on these models, two maps were created: a fire risk map where the risk of fire is associated to the density of fires and a fire danger map where the danger of fire is associated to the area burned. Lastly, the data was transformed into incidence data, binary data indicating the presence of fire, and fire risk maps were built based on transition probabilities (IvIarkov model) and the hazard function of the survival models (discrete Weibull model). In all models were considered several covariates. Regarding the results obtained, the models allow to identify areas where more fires occurred as well as the largest fires, however the implemented models based on space-time point processes underestimates the density and size of fires that actually occurred. The inference on these mo deIs were carried using Bayesian hierarchical models, the SPDE approach and the IN LA approach, except in the discrete Weibull model where the inference was carried via the simulation methods IvICIvIC through WinBUGS. |
|---|