Active fire segmentation in satellite images using machine learning: cross-sensor adaptation strategies

Abstract

Active fire segmentation on satellite imagery is a fundamental activity for environmental monitoring. This task is commonly performed through thresholding methods based on equations designed for specific sensors. Our research focuses on active fire segmentation using machine learning techniques applied to satellite imagery, thereby eliminating the need for manual threshold definition. More specifically, our study aims at analyzing different strategies and deep learning models to exploit similarities between distinct sensors, reducing the amount of data required to train a model for a given satellite. The Landsat-8 and Sentinel-2 satellites were chosen as the objects of study due to the similarity in their observed spectra. In the first part of this research, we evaluated various transfer learning strategies to adapt a deep network trained on Landsat-8 images to Sentinel-2, using a limited number of labeled samples to achieve superior results compared to traditional thresholding algorithms. Additionally, we analyzed the impact of the main spectral bands used for active fire segmentation. In the second part, we examined the performance of different architectures when adapted between the two satellites, as well as evaluating the impacts of issues commonly encountered in Sentinel-2 imagery on both the models and the thresholding methods. In the third part, we investigated the potential of different architectures, including foundation models, for the development of a new active fire segmentation approach. To this end, we propose a hybrid solution that combines a deep network with a branch dedicated to the application of automatically learned indices. Besides being inherently interpretable, the index branch presents low computational cost and a reduced number of parameters, acting in a complementary manner to maximize the robustness and generalization capacity of the detection.