Day 4

Detailed paper information

Back to list

Paper title Global Active Fire Detection – Towards a SAR-enabled Multi-Sensor Global Monitoring System
  1. Deniz Gaye Denizoglu Techische Universität München Speaker
  2. Gabriel Dax Technische Universität München
  3. Srilakshmi Nagarajan Technical University of Munich
  4. Ning Zhang Technical University of Munich, Changchun Institute of Optics, Fine Mechanics and Physics, Changchun; Chinese Academy of Sciences; University of Chinese Academy of Sciences Beijing;
  5. Martin Werner Technical University of Munich
Form of presentation Poster
  • C1. AI and Data Analytics
    • C1.04 AI4EO applications for Land and Water
Abstract text Active fire detection for environmental monitoring is a very important task that can significantly be supported by satellite image analysis. Active fires need to be detected not only for fire fighting in settled areas, but also for finding fires in the wilderness, which is only possible from satellites global coverage.

Classically, active fire detection is based on multispectral signatures of fire on a per-pixel basis, sometimes including statistics of the surroundings. Such classical methods are fast, easy to apply and surprisingly powerful both in detecting and dissecting active fires. Following related work from Pereira [1], our work is based on fire detection algorithms from Schroeder [2], Kumar-Roy [3], and Murphy [4] combined with methodological inspiration form modern deep learning.

Recent work on fire detection has been given in [5]. The authors use fire perimeter data from the California Fire Perimeter Dataset (CALFIRE ) to create a multi-satellite collection of training data for fire segmentation. While using all satellites is an extremely interesting aspect of this work, the training data generation process is tailored to known fires in a small region of the world only and cannot safely distinguish active fires from burnt areas.

Pereira et al. use a completely orthogonal approach on a global scale [1]. They apply three different, simple, explainable, and well-known active fire detection methods on Landsat multispectral images to derive global active fire detection training data and train some basic U-Net models on this data successfully. In contrast to the first paper, however, they rely on a single satellite system.
Both papers are excellent contributions to the problem of fire detection from Earth observation data. A combination of their methodology, however, combined with a more advanced data management and analysis pipeline is promising.
In this project, we work towards closing the gap by using the Landsat data together with the given deterministic fire detection methods and fit minimalistic deep neural networks to reproduce the exact same of multispectral detections on Sentinel-2 data. Thereby, the traditional active fire detection models designed for Landsat instruments are safely transformed to input data from ESA mission Sentinel 2.
Based on this, we extend the work to integrate SAR data from Sentinel 1 and various methodologies of data preparation and fusion. For example, we apply a data preparation scheme based on a genetic algorithm for finding good representations of the whole multispectral information for this task [6] and we apply an automated model fusion technique we previously applied to building instance classification with success [7].

The outcome of this project is a methodology to derive global active fire datasets, which might suffer from errors of the underlying deterministic methodology and the transformation process, but which allow for global fire monitoring, which is of high interest in the context of climate and deforestation analysis together with baseline models both from simple data mining and deep learning regimes.
In the poster, we want to present our early results giving hints on the baseline performance of all steps, which we are going to improve during the course of this master thesis research project.


[1] G. H. Almeida Pereira, A. M. Fusioka, B. T. Nassu and R. Minetto, "Active fire detection in Landsat-8 imagery: A large-scale dataset and a deep-learning study," ISPRS Journal of Photogrammetry and Remote Sensing, vol. 178, p. 171–186, 2021.
[2] W. Schroeder, P. Oliva, L. Giglio, B. Quayle, E. Lorenz and F. Morelli, "Active fire detection using Landsat-8/OLI data," Remote Sensing of Environment, vol. 185, p. 210–220, 2016.
[3] S. S. Kumar and D. P. Roy, "Global operational land imager Landsat-8 reflectance-based active fire detection algorithm," International Journal of Digital Earth, vol. 11, no. 2, p. 154–178, 2018.
[4] S. W. Murphy, C. R. Souza Filho, R. Wright, G. Sabatino and R. Correa Pabon, "HOTMAP: Global hot target detection at moderate spatial resolution," Remote Sensing of Environment, vol. 177, p. 78–88, 2016.
[5] D. Rashkovetsky, F. Mauracher, M. Langer and M. Schmitt, "Wildfire Detection From Multisensor Satellite Imagery Using Deep Semantic Segmentation," IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 14, p. 7001–7016, 2021.
[6] G. Dax, M. Laass and M. Werner, "Genetic Algorithm for Improved Transfer Learning Through Bagging Color-Adjusted Models," 2021, p. 2612–2615.
[7] E. J. Hoffmann, Y. Wang, M. Werner, J. Kang and X. X. Zhu, "Model Fusion for Building Type Classification from Aerial and Street View Images," Remote Sensing, vol. 11, no. 11, 2019.