Mapeamento automático de queimadas no bioma Cerrado utilizando sensores orbitais

Abstract

The objective of this dissertation was to develop an automatic algorithm to map wildfires in the cerrado biome using orbital sensors. To do this, four articles were developed. The first analyzed eight spectral indexes, commonly used to map wildfires in Landsat images, evaluated by means of the M separability index. The study was conducted at a Conservation Unit mosaic in northern Minas Gerais, Brazil. The NBR2 index obtained greater separability, with the value of M of 2.5, considered the most indicated to map wildfires in this region by Landsat images. In the second article, an automatic algorithm was developed to map wildfires in Landsat-8 images for the same region as studied in article 1. To do this, a multi-temporal composite of six Landsat images, with date of the critical wildfire period of 2015, based on the pixel choice of the lowest NBR2 index value. The wildfire samples were collected by active hotspots, and used to train the Support Vector Machine (SVM-OC) single class classifier. Three kernel and different combinations of SVM-OC parameters were evaluated in order to verify which were most adequate in mapping wildfires. The radial kernel presented higher accuracy, with kappa index of 0.98. The results showed that 13% of the mapped burnt area were scars with no active hotspots. The third article evaluated four multi-temporal composite techniques, using PROBA-V, images regarding the capacity of discriminating burnt areas and the presence of cloud shadows. The technique that uses the second lowest reflectance value of the near infrared channel (NIR) obtained separability little lower than the technique of the lowest reflectance value of the NIR (M indexes of 1.3 and 1.4, respectively). However, it presented images with less cloud shadows, being considered the most appropriate for mapping wildfires in PROBA-V images. In the fourth article, an algorithm was developed to map wildfires in the PROBA-V multi-temporal composites, validated with wildfire maps in Landsat images (reference), comparing the results with the MODIS MCD64A1 product. The PROBA-V product presented total omission of 30%, while MCD64A1 presented 34%. The commission errors were smaller for MCD64A1 when compared to the PROBA-V (15% and 22%, respectively). PROBA-V obtained the best results in all analyzed scenarios, analyzing the wildfire correlation in a 10x10 km grid, calculated by means of the Kendall coefficient, showing that the developed algorithm can improve the estimates of burnt areas in the cerrado biome.