Multi-temporal abrupt change estimation on Landsat time series imagery: an application to analyze burn severity in La Primavera, Mexico
share
We analyze 14-years NDVI and dNBR time series of Landsat-7 imagery for monitoring long-term burn severity on an Area of Protection of Flora and Fauna known as La Primavera. We propose a burned area mapping approach that does not require prior knowledge of the fire's date. We begin by applying a statistical algorithm (BFAST) to estimate abrupt changes in the NDVI's trend. The estimated abrupt changes are used as reference dates (plausible fire dates) from which dNBR is calculated following a typical pre-post fire assessment. These dNBR values allow us to determine burned areas and quantify an apparent severity or a sudden regrowth. Unlike most burned area mapping methods ours depends effectively only on one tuning parameter, h; based on a set of simulations we provide rules to select this parameter. Our simulations also show that when one abrupt change is present in the NDVI's trend, BFAST's performance is appropriate even when 20 probability of estimating correctly two abrupt changes deteriorates as the percentage of missing values increases, BFAST's conditional underestimation will cease when the distance between abrupt changes is greater or equal than the sample size times h and we did not find overestimation. We use a RapidEye burned area map to validate our procedure; we show empirically that our maps can achieve up to 83 series in which 47 to 49 annual burn severity maps that allow us to establish that La Primavera has undergone to a series of statistically significant vegetation changes, plausibly attributed to fire, almost in its entirety throughout the studied period. Despite the latter, we found that most of La Primavera's burned areas have suffered from low severity. R code to implement our approach complements this paper.
READ FULL TEXT
