MACHINE LEARNING IN THE MODELING OF FOREST FIRES IN THE STATE OF ESPÍRITO SANTO
Name: RONIE SILVA JUVANHOL
Type: PhD thesis
Publication date: 19/12/2017
Advisor:
Name |
Role |
|---|---|
| ALEXANDRE ROSA DOS SANTOS | Co-advisor * |
| NILTON CESAR FIEDLER | Advisor * |
Examining board:
| Name |
Role |
|---|---|
| ALEXANDRE ROSA DOS SANTOS | Co advisor * |
| JOSÉ EDUARDO MACEDO PEZZOPANE | Internal Alternate * |
| NILTON CESAR FIEDLER | Advisor * |
| TELMA MACHADO DE OLIVEIRA PELUZIO | External Examiner * |
| WELLINGTON BETENCURTE DA SILVA | External Examiner * |
Summary: The main problem encountered when applying geographic information systems and remote sensing techniques for the prediction of forest fires is the necessity to integrate different data sources. The methods applied are usually based on regression techniques or on coefficients that depend on expert knowledge. The objective of this study was to test the capacity of the classification and regression tree (CART) to assess the regional fire risk. The CART analysis is a non-parametric statistical technique that generates decision rules in the form of a binary tree, for a classification or regression process. The MCD45A1 product of burn area, relative to 16-year (2000-2015) was used to obtain a fire occurrence map, from the center points of the grid cell, using a kernel density approach. The resulting map was then used as input response variable for the CART analysis with fire influence variables used as predictors. A total of 12 predictors were determined from several databases, covering environmental physical and socioeconomic aspects. The rules induced by the regression process allowed the definition of different risk levels, expressed in 35 management units, used to produce a fire prediction map. According to the results, the Northeast Region, sweet river and Southeast represent the major risk areas in the state (South Coast). The results of the regression process (r = 0.94 and r² = 0.88), the capability of the CART algorithm analysis to highlight the hierarchical relationships between the predictor variables and the easy interpretability of the decision rules represent a possible tool to better approaching the problem of assessing and representing forest fires.
Keywords: Non-parametric statistics, Kernel density, CART algorithm, Decision rules, Fire prediction map.
