Spectral angle indices for burned area detection in Chile using Sentinel-2 data and a Random Forest classifier

Abstract:

Every year, fires are more frequent and intense due to weather and vegetation conditions. In January 2017, 114 active fires burned throughout Chile at the same time. These fires spread quickly due to the high temperatures, strong dry winds, and low vegetation water content. The fire events burned more than 570,000 ha, from which 20% of the area was endangered native forest. Almost half of the area burned in Chile 2017 occurred in the Region of Maule.

This study aimed to develop and implement an algorithm for burned area (BA) classification using Sentinel-2 data at 20 meters resolution on Google Earth Engine (GEE) computing platform. GEE allows access to an extensive database of various satellite imagery datasets and a powerful ability of data processing. Because of frequent cloud cover in the region, we computed temporal composites to ensure the analysis of every pixel. The pre and post-fire composites were generated by minimizing the Normalized Burn Ratio spectral index (NBR). This minimization selects burned pixels while dismissing most clouds, cloud shadows, and snow.

We trained a random forest (RF) classifier to account for all the land cover variability and ensure accurate classification of BA. We included as input variables the Sentinel-2 bands, the NDVI, NBR, SAVI, MIRBI, and three spectral angle indices (Shortwave angle slope index, Shortwave angle normalized index and the angle at NIR). The classifier was trained from a random point sample generated using as reference the official fire perimeters. We ran the RF classifier with and without spectral angle indexes (SAI) to account for the effect of these indices in the classification. Spectral indices showed a high value of importance index computed by the RF algorithm. According to the visual assessment, both classifications of BA are more accurate than the perimeters created by the Chilean National Forest Corporation, which overestimates the area burnt because it includes as burnt the inner unburned areas and, it omits some small burned areas. Comparing both BA classifications, with and without SAIs, we observed differences in the spatial distribution of the errors. The BA classification with SAIs offered higher accuracies, with errors of omission and commission of 8 and 16%, respectively.