Abstract

TECUAPETLA-GOMEZ, Inder; VILLAMIL-CORTEZ, Gabriela  and  CRUZ-LOPEZ, María Isabel. Statistical Estimate of Burned Areas in La Primavera (Mexico) from 2003 to 2016 Using Time Series of Landsat-7 Images. Invest. Geog [online]. 2021, n.106, e60418.  Epub June 06, 2022. ISSN 2448-7279.  https://doi.org/10.14350/rig.60418.

The La Primavera Wildlife Protection Area (APFFLP, in Spanish) is located west of Guadalajara, the third largest city of Mexico. This natural area harbors four types of vegetation: oak forest (Quercus), oak-pine forest (Quercus-Pinus), pine-oak forest (Pinus-Quercus), and deciduous tropical forest, as well as three plant communities - riparian, rocky, and ruderal - that grow within the vegetation types just mentioned. The natural richness of La Primavera is potentially threatened due to forest fires that occur at least once a year.

This text reports the results of an analysis to determine semi-automated areas burned in La Primavera during the period 2003 to 2016 using time series of Landsat-7 images of the Normalized Difference Vegetation Index and the Normalized Burn Ratio, NDVI and NBR, respectively.

To determine burned areas in the APFFLP, we first estimated the dates of abrupt changes in the trend of the NDVI time series using the Breaks For Additive Season and Trend (BFAST) algorithm. In a first instance, this algorithm allows estimating the trend, seasonal or temporal component, and the corresponding errors of each NDVI time series. Moreover, by combining dynamic programming, a statistical model selection criterion, and a statistical hypothesis test, BFAST allows estimating abrupt shifts in the trend and seasonal component of the NDVI; the algorithm returns the dates of these abrupt changes along with their respective confidence interval. This study only considered very marked changes in the NDVI trend, which were deemed plausible burn dates.

Because abrupt changes in the NDVI structure may have different root causes, such as deforestation, it was necessary to determine their association with level changes in a spectral index used to determine burned areas. To this end, we used NBR values in a typical pre-post-burn analysis. That is, having a date of abrupt change in the NDVI trend as reference, we calculated the difference in NBR values between the previous date (one year apart, to reduce any potential interference from phenology or illumination effects) and the date immediately after the estimated abrupt change. Based on studies that have used of the difference in NBR in areas with a similar vegetation cover as La Primavera, we used the values of this index to determine proxy-severity levels in the burned areas determined by our method.

To note, the method described herein is semi-automated and independent from the expertise of the researcher to determine burned areas. This is particularly useful for monitoring areas with no fire statistics available. Our map on the La Primavera burned area for 2008 illustrates that this method allows the detection of burned areas with no prior knowledge of the incident.

BFAST, and therefore our method for the identification of burned areas, depends on a bandwidth parameter used in the statistical test to determine abrupt changes. The specific features of this parameter were widely discussed for this text. Since the Landsat-7 image collection for the study period had a high proportion of missing data, we discussed various information-filling techniques and studied their impact on the actual estimation of burned areas.

The frequency of burned areas estimated with our method is in line with the results of other studies carried out in situ for the period 2003-2016 by other research groups. According to our analysis, the severity of most of these fires is low. To validate our estimate of burned area, we used high-resolution images (RapidEye), calculating matrices of omission and commission errors. From them, it follows that, regardless of the interpolation method used to fill missing data in the images and the value of the BFAST bandwidth parameter, the accuracy of our map of burned area ranges from 70% (with low-quality data) to 92% (with moderate-quality data). This note is supplemented with open software, available in the GitHub website.

Keywords : burned areas; NDVI; NBR; time series; BFAST; Landsat-7; La Primavera.

        · abstract in Spanish     · text in Spanish     · Spanish ( pdf )