On the potential of BFAST for monitoring burned areas using multi-temporal Landsat-7 images

TL;DR

This paper presents a semi-automatic method using BFAST on NDVI time series to detect burned areas and assess severity without prior fire date knowledge, validated with Landsat-7 data over Mexico.

Contribution

It introduces a novel approach combining BFAST change detection with burn severity assessment that requires minimal prior information and handles data gaps effectively.

Findings

Achieved 92% accuracy in burn area mapping.

Effectively detected burn severity over multiple wildfire events.

Validated method with high-resolution RapidEye data.

Abstract

In this paper, we propose a semi-automatic approach to map burned areas and assess burn severity that does not require prior knowledge of the fire date. First, we apply BFAST to NDVI time series and estimate statistically abrupt changes in NDVI trends. These estimated changes are then used as plausible fire dates to calculate dNBR following a typical pre-post fire assessment. In addition to its statistical guarantees, this method depends only on a tuning parameter (the bandwidth of the test statistic for changes). This method was applied to Landsat-7 images taken over La Primavera Flora and Fauna Protection Area, in Jalisco, Mexico, from 2003 to 2016. We evaluated BFAST's ability to estimate vegetation changes based on time series with significant observation gaps. We discussed burn severity maps associated with massive wildfires (2005 and 2012) and another with smaller dimensions (2008) that might have been excluded from official records. We validated our 2012 burned area map against a high resolution burned area map obtained from RapidEye images; in zones with moderate data quality, the overall accuracy of our map is 92%.