Burned Area Mapping in the Brazilian Savanna Using a One-Class Support Vector Machine Trained by Active Fires

Duarte Oom,José Pereira,Luis de Carvalho,Renata Libonati,Fausto Machado-Silva,Alberto Setzer,Allan Pereira,Fabiano Morelli

doi:10.3390/rs9111161

Abstract

We used the Visible Infrared Imaging Radiometer Suite (VIIRS) active fire data (375 m spatial resolution) to automatically extract multispectral samples and train a One-Class Support Vector Machine for burned area mapping, and applied the resulting classification algorithm to 300-m spatial resolution imagery from the Project for On-Board Autonomy-Vegetation (PROBA-V). The active fire data were screened to prevent extraction of unrepresentative burned area samples and combined with surface reflectance bi-weekly composites to produce burned area maps. The procedure was applied over the Brazilian Cerrado savanna, validated with reference maps obtained from Landsat images and compared with the Collection 6 Moderate Resolution Imaging Spectrometer (MODIS) Burned Area product (MCD64A1) Results show that the algorithm developed improved the detection of small-sized scars and displayed results more similar to the reference data than MCD64A1. Unlike active fire-based region growing algorithms, the proposed approach allows for the detection and mapping of burn scars without active fires, thus eliminating a potential source of omission error. The burned area mapping approach presented here should facilitate the development of operational-automated burned area algorithms, and is very straightforward for implementation with other sensors.

Highlights

Vegetation burning is a global-scale process that affects the global distribution and structure of vegetation, major biogeochemical cycles, and the climate system [1]
Project for On-Board Autonomy-Vegetation (PROBA-V) near-infrared reflectance (NIR) daily reflectance values, with spatial resolution of 300 m, are georeferenced based on coordinate information contained in the metadata, rejecting pixels: (1) containing solar zenith angles greater than 60◦ and/or viewing zenith angles of NIR channel greater than 40◦; (2) classified as cloudy in the PROBA-V Quality assurance layers; (3) containing low radiometric quality; (4) containing reflectance values higher than 0.5
The AQM-PROBA is based on the spectral space of the NIR T2 and NIR T1-T2 values in PROBA-V images, with samples collected by Visible Infrared Imaging Radiometer Suite (VIIRS) active fire, which is used to train the OC-Support Vector Machine (SVM) classification model

Summary

Introduction

Vegetation burning is a global-scale process that affects the global distribution and structure of vegetation, major biogeochemical cycles, and the climate system [1]. Wildfires are a socio-natural hazard that annually affect millions of hectares of forests, woodlands, and other vegetation, endangering human populations, and causing substantial economic losses, both in terms of assets destroyed and in the form of prevention and suppression costs [2]. The Brazilian savanna has been increasingly affected by deforestation due to cropland and pasture expansion, increasing and altering the natural fire regime in the region [3,4]. Attempts to characterize these anthropogenic impacts presuppose understanding of spatial and temporal fire patterns [1]. Despite the high frequency of human induced-fire and significant disturbance caused to the Cerrado biome, fire dynamics are not yet well characterized

Objectives

Methods

Results

Discussion

Conclusion