The year 2019 was especially bad year for forest fires in the Amazon region, and there are reports that are revealing a rise in deforestation of nearly 90% between August 2019 and May 2020, compared to the period of August 2018 and July 2019. With the intention of estimating the actual extent of landcover damage done by a detected fire, Sentinel 2 images are introduced in an object-based classification environment (eCognition) in order to develop a classification procedure for burnt area mapping.
The aim of this study is the generation of a burn severity map for the assessment of the areas affected by wildfires in the Amazon region in August of 2019, particularly we looked at an area in the region of Borochi, Bolivia.
Satellite images from Sentinel 2B were the primary data used in this study. These images were acquired from the Copernicus Open Access Hub. Two images corresponding to the AOI for the dates: 28 of July (pre-fire) and 22 of August (post-fire). These are Level 2A products from Copernicus which means Top of Atmosphere correction has been applied already.
Other validation products were also acquired from various sources including:
The downloaded satellite images were preprocessed wherein a composite was created in ArcGIS Pro. The new image contains all the needed bands in a single image.
Index calculations, segmentation and initial classification were done in eCognition. The Normalized Burn Ratio (NBR) was used, as it was designed to highlight burned areas and estimate burn severity. It uses near-infrared (NIR) and shortwave-infrared (SWIR) wavelengths. Healthy vegetation before the fire has very high NIR reflectance and a low SWIR response. In contrast, recently burned areas have a low reflectance in the NIR and high reflectance in the SWIR band. More information about the NBR can be found here. The NBR is calculated for images before the fire (pre-fire NBR) and for images after the fire (post-fire NBR) and the post-fire image is subtracted from the pre-fire image to create the differenced (or delta) NBR (dNBR) image. dNBR can be used for burn severity assessment, as areas with higher dNBR values indicate more severe damage whereas areas with negative dNBR values might show increased vegetation productivity. dNBR can be classified according to burn severity ranges proposed by the United States Geological Survey (USGS). (Source: UN spider.org)
The validation process was done in ArcGIS Pro using available tools for Accuracy Assessment. Random sampling was used to generate sample points from the EMS and MODIS burnt area shapefiles.
Confusion matrix and Kappa statistics were also generated to assess the accuracy of the classification.
The resulting classifications were presented in different maps. First, the burnt area extent can be seen in the map below.
The total burnt area were estimated for two time periods since we noticed that there were already burnt scars from the first image capture. To estimate the burnt forested area, we took the difference between the classified forest and the burnt area. The numbers show that about 300 ha. of forest area were burnt in the period of one month, while a large part of the burnt area was classified as non-forest.
Using the USGS Severity threshold values, the following severity levels were derived for the burnt area.
For the validation, the confusion matrix values and Kappa value are as follows:
The OBIA approach applied to Sentinel 2 imagery has proven to be successful in generating accurate estimations of burnt area and severity in the Bolivian Amazon region. This approach can be used as an alternative and effective procedure for emergency mapping of wildfire damage. The present study detected a total of 3,4% less burnt area than the EMS product for the same area, which is not a significant difference, and therefore supports the accuracy of the results. However, the presented procedure needs to be tested in different areas, seasons and using additional validation products to further assess its effectivity in detecting burnt areas for the Amazon forests.
*This short case study was accomplished as a collective effort of our group during the GEOBIA Summer School 2020 and was presented in the AGIT Conference 2020 in Salzburg, Austria.
Group Members:
Jenny Valiente
Maliha Malek
Rabindra Adhikari
Leonard Luz
Sayana de Gorostizaga
