Diagnosis and prediction of forest fires using artificial neural networks
Name: FELIPE PATRICIO DAS NEVES
Publication date: 24/02/2022
Advisor:
Name |
Role |
|---|---|
| ALEXANDRE ROSA DOS SANTOS | Co-advisor * |
| NILTON CESAR FIEDLER | Advisor * |
Examining board:
| Name |
Role |
|---|---|
| ELAINE CRISTINA GOMES DA SILVA | External Examiner * |
| FLAVIO CIPRIANO DE ASSIS DO CARMO | External Examiner * |
| NILTON CESAR FIEDLER | Advisor * |
| SÁULO BOLDRINI GONÇALVES | External Examiner * |
Summary: In recent years forest fires have been increasingly frequent in several regions of the world. As a result, several institutions have sought strategies capable of mitigating its consequences, as they can cause serious damage to society, amid limited financial, human and material resources. In this context, this research aimed to carry out a diagnosis of the historical series of services provided by the Military Fire Brigade to forest fires in the state of Espírito Santo, from 2010 to 2019. Then, it analyzed details related to firefighting actions, such as time of commitment of teams, displacements, impacted vegetation, periods of demand. In addition, a hybrid fire map model was proposed, gathering data from the Integrated Operational and Social Defense Center (CIODES) and the National Institute for Space Research (INPE); Artificial neural networks (artificial intelligence) were modeled to predict new outbreaks in the north and northwest of the state, based on the historical series of fires recorded by CIODES and meteorological data from stations of the National Institute of Meteorology (INMET). For statistical analysis, the attendance data were submitted to the analysis of variance test, and, when F was significant, means were submitted to the Tukey test at a significance level of 5%. The software ArcGIS 10.8 and QGIS 3.16, and the Statistical Package for the Social Sciences (SPSS), from the International Business Machines Corporation (IBM), were used in the creation of maps and spatial analysis, in the construction of artificial neural networks and prediction of forest fires. . The years 2015 and 2019 stood out as critical periods; August to October, and January to March (winter and summer) with higher annual incidence; as well as between Fridays and Sundays. It was also found that forest fires were more frequent in the administrative micro-regions Metropolitana, Rio Doce, Midwest, Northeast and Central Sul. Both CBMES and INPE records showed similar trends, differing in absolute numbers due to the methodology used by Organs responsible bodies. In addition, there was a general increase in the monthly averages of commitment time and displacement of the Corporation s vehicles in 2018 and 2019; the most impacted vegetation was non-native (66.91%), in addition to native (17.88%), and undergrowth/wasteland (15.21%). The hybrid map using data from CIODES and INPE proved to be relevant, capable of helping responsible managers in planning future
actions. Finally, it was possible to model artificial neural networks for the 05 most punished municipalities in the north and northwest of the state and validate them, having their performance metrics determined, showing generally good accuracy and precision. Regarding the importance of variables, wind speed stood out in the Linhares and Nova Venécia networks, while relative air humidity for the municipalities of Colatina and Aracruz, and temperature for the São Mateus network. When subjected to hypothetical situations, they were able to predict foci for each condition and location proposed in the study. In this way, the research obtained important technical information, capable of supporting strategic decisions of competent managers, seeking to optimize the use of public resources in actions aimed at this type of disaster.
Keywords: Forest Protection, Fire prevention and fighting, Geotechnologies, Artificial Neural Network.
