New techniques for old fires: Using deep learning to augment fire maps from the early satellite era

Picture
Figure 2. Results Overview
Picture
Figure 3. Results Detail A. MSS images and model classifications of a newly detected fire, along with a timeline of each model classification for a single pixel.
Picture
Figure 7. Results Detail E. MSS images and model classifications of a fire detected by our methods and listed in the existing databases, along with a timeline of each model classification for a single pixel.
Background

Fire is a fundamental ecological process that shapes forest structure, carbon storage, and habitat over decades. To understand present-day forests and manage them effectively, we need accurate records of where and when past fires occurred. However, many fires from the 1970s are missing or poorly mapped in existing databases because they predate modern satellite systems. Imagery from the early Landsat era is difficult to interpret due to coarse resolution, limited spectral information, and sparse observations. As a result, a critical decade of fire history remains incomplete, limiting our ability to assess long-term fire regimes and their ecological consequences.

Approach

In this paper, we developed a deep learning–based approach to recover fire history from early satellite imagery. We used imagery from the Landsat Multispectral Scanner, the only global Earth observation system operating in the 1970s, and trained a neural network to recognize burn scars despite the data’s limitations. We first carefully preprocessed thousands of images to reduce noise and highlight burn-related signals. We then trained a convolutional neural network to classify burned and unburned areas at the pixel level. By combining these classifications through time, we produced annual fire maps and filtered out false detections, resulting in a consistent and spatially detailed record of historic fires.

Key Findings
​
Our analysis substantially expanded the known fire history of Quebec’s forests:
  • Major increase in mapped fire area: We identified approximately 3,500 km² of burned area that was not included in any existing fire database.
  • Significant revision of historical records: These newly detected fires represent about a 35% increase in the total known burned area between 1973 and 1982 in forest-dominated regions of Quebec.
  • Strong overall accuracy: The final burned-area maps achieved very high overall accuracy, while revealing complementary strengths and weaknesses relative to existing fire databases.
  • Improved spatial detail: Our maps often excluded unburned islands and refined fire boundaries that were previously represented as coarse or generalized polygons.

Impact

This work demonstrates that modern deep learning methods can unlock valuable information from early satellite archives. By extending reliable fire maps back into the 1970s, the lab provides a more complete picture of historical fire regimes, strengthening studies of carbon cycling, forest recovery, and climate–fire feedbacks. The approach is scalable and adaptable, offering a pathway to extend disturbance and land-cover records across Canada and other boreal regions, and to support better-informed forest and climate policy decisions.

​

Resources

Published Paper : Boothman* R, Cardille JA. New techniques for old fires: Using deep learning to augment fire maps from the early satellite era. Frontiers in Environmental Science. 2022 August 17. DOI: https://doi.org/10.3389/fenvs.2022.914493