Mapping Burned Areas with Optical and Sentinel-1 Synthetic Aperture Radar Data

This work focuses on monitoring wildfires using remote sensing and time series change detection of multi-spectral and Synthetic Aperture Radar (SAR) data. Remote sensing can offer detailed information on fire conditions and risks paired with frequent revisit times. Multispectral satellites, like Landsat and Sentinel-2 (S-2), data have been, and still are, widely used in Earth Observation thanks to the wide range of wavelengths and the frequent rate of observation. On the other hand, Sentinel-1 (S-1) SAR data offers the advantage of independence from solar illumination and weather conditions. This work focuses on generating and classifying time series of Landsat and Sentinel-2 data into burned/unburned areas using a random forest (RF) algorithm. Python and Google Earth Engine (GEE) tools were developed to automate the extraction of fire reference perimeters from multi-spectral images and merging the results to form a burned area dataset, while providing support to the user. S-2 burned area perimeters over the Sahel (Africa) and Amazon (South America) regions were used to assess the sensitivity to burned surfaces of S-1 SAR data pre- and post-fire backscatter in the VV and VH polarizations. Statistical tests and data visualization were used to assess the change in intensity of the backscatter signal (γ0) after a fire event. The changes in the distribution of γ0 values were found significant for the majority of the test cases, and more visible in the VH polarization mode. Classification tests were conducted using the S-1 backscatter values and dedicated radar indices to identify the occurrence of fires, considering different time configurations. Longer temporal baselines of S-1 acquisitions were found to produce more accurate results in detection in high fire activity regions. Although S-1 SAR shows promise in wildfire detection, challenges remain in interpreting radar returns. Further research could explore other polarization methods and longer temporal baselines to enhance accuracy in different biomes.