Optimization of a Sentinel-2 burned area processor by integration of deep learning based smoke segmentation

Abstract

Due to the increased occurrence and intensity of wildfires around the globe, the development of sophisticated segmentation techniques in satellite images is crucial for effective hazard management. In order to improve the wildfire burned area processor integrated at the German Aerospace Center (DLR), a smoke segmentation model is proposed to identify smoke pixels in Sentinel-2 images. This will allow for more accurate burned area mapping and will provide useful auxiliary information to aid those combating wildfires and managing the aftermaths. The proposed smoke segmentation model is based off of U-Net, an encoder-decoder convolutional neural network. A dataset is created of scenes captured over different locations around the world in order to create a robust sample containing as many variations as possible. Consequently, the images are divided into training, validation and testing groups and then are tiled into smaller segments with the size of 256 x 256 pixels. Ground truth masks are generated through an automated thresholding process with manual corrections as needed. Using these pairs of images and masks, fifteen trials were conducted to evaluate the following aspects: the inclusion of data augmentation, the use of class weighting, the best optimizer, and the optimal combination of spectral bands input to the model. The results showed that the use of data augmentation and class weighting is beneficial for creating a robust segmentation model. Additionally, the best optimizer was found to be Adam, and the band combination that produced the best results included the following eight bands: coastal aerosol, blue, green, red, near infrared (NIR), short wavelength infrared (SWIR) – cirrus, SWIR1 and SWIR 2. The final selected model has a Cohen’s kappa coefficient of 0.92, an accuracy of 0.98 and an F1-score of 0.99 for smoke pixels and 0.93 for non-smoke pixels, proving that smoke can be segmented well using deep learning in Sentinel-2 images. Thus, this model was successfully implemented in the S2-BAP-cnn burned area processor at DLR for operational use.