Active Fire Segmentation: A Transfer Learning Study From Landsat-8 to Sentinel-2

Active fire segmentation through satellite imagery is fundamental to prevention and damage prediction. Algorithms to address this problem usually rely on sensor-specific thresholds, empirically chosen based on a few image samples, and thus, are susceptible to many errors. Deep learning algorithms automatically extract information through various levels of abstraction, avoiding human intervention for feature engineering. However, training a deep network from scratch requires massive labeled data, and efforts in this direction have already been made for the Landsat-8 satellite. For Sentinel-2, another important satellite that has band wavelength similarities with Landsat-8, there is a limited number of such initiatives, with few studies even concerning hand-crafted (traditional) algorithms. In this context, we explored in this article the transfer of knowledge for active fire segmentation from Landsat-8 to Sentinel-2, avoiding the need of a vast amount of labeled data from Sentinel-2, and also reducing the computational resources for training. We also compiled a benchmark containing 12 584 image patches extracted from 26 Sentinel-2 images from around the globe, along with manually annotated fire pixels, to assess the algorithm's response compared to a human specialist. In a series of transfer learning experiments by using the U-Net architecture, we showed that even a single labeled image from Sentinel-2 for training was sufficient to allow achieving an
$F$
-score metric of 84.9% when Landsat-8 knowledge is transferred and further fine-tuned in the machine learning process, while the best hand-crafted algorithm designed for Sentinel-2 achieved an
$F$
-score of 75.8% in the segmentation task.