CanLaBS v2 is an update to the Canada Landsat Burned Severity (CanLaBS) data product, available at https://doi.org/10.23687/b1f61b7e-4ba6-4244-bc79-c1174f2f92cd, builds upon the methodology originally described in Guindon et al. (2021), entitled “Trends in wildfire burn severity across Canada, 1985 to 2015” and published in the Canadian Journal of Forest Research (https://doi.org/10.1139/cjfr-2020-0353). CanLaBS v2 introduces several important improvements to input data sources, temporal coverage, and modeling approaches. **1. Key Updates in CanLaBS v2** **1.1 Transition to Landsat Collection 2** All Landsat inputs used to derive burn severity metrics have been updated from Landsat Collection 1 to Landsat Collection 2 (Earth Resources Observation and Science (EROS) Center, 2020a, 2020b, 2020c). Landsat Collection 2 provides improved radiometric calibration, refined atmospheric correction, and enhanced geometric accuracy, resulting in greater temporal consistency and more reliable spectral change detection across sensors and years. **1.2. Expanded fire perimeter coverage (NBAC 1986–2024)** The updated product now covers all fire perimeters included in the National Burn Area Composite (NBAC; Skakun et al., 2022) from 1986 to 2024. This substantially extends the temporal range of the dataset relative to the original release and ensures consistency with the most up-to-date national fire perimeter record used in Canada-wide disturbance analyses. **1.3. Improved random forest model for salvage logging detection** Salvage logging detection has been updated using an improved random forest (RF) classification model trained on 3614 photo-interpreted reference points. The model uses a refined set of spectral predictors derived from Landsat imagery, including pre- and post-fire band 3, post-fire bands 4, 5 and 7 (according to the Landsat 7 nomenclature), inter-annual spectral differences (ΔB3, ΔB4, ΔB5), and pre- and post-fire Normalized Difference Vegetation Index (NDVI). Model performance was evaluated using a train-test split (80%, 20%, respectively). This analysis revealed an overall accuracy of 90.6% and Cohen’s kappa of 0.87 (Table 1). Some confusion occurred between low-vegetation fires and salvage logging (the primary class of interest), but overall performance was strong, with 95.49% precision, 75.6% recall, and an F1-score of 84.39%. **Table 1.** Test set confusion matrix of the salvage logging detection random forest model. | Observed / Predicted | No Fire | Fire | Low vegetation fire | Salvage logging | |------------------------------|-----------:|----------:|--------------------------:|-----------------------:| | No Fire | 160 | 1 | 0 | 1 | | Fire | 4 | 148 | 7 | 2 | | Low vegetation fire | 0 | 9 | 220 | 3 | | Salvage logging | 4 | 8 | 29 | 127 | **1.4. Revised gapfilling strategy** As in the original product, gapfilling of pre-fire Landsat data is retained to ensure complete characterization of pre-disturbance conditions. However, post-fire Landsat gapfilling is no longer applied in this version. This results in some missing data but avoids the introduction of uncertainty associated with radiometric regression-based gapfilling. A total of 6.9% of all NBAC burnt pixels are missing data. This proportion decreased over time due to improved Landsat data coverage, from 12.7% for fires before 2000 (pre-Landsat 7) to 2.59% for fires after 2012 (post-Landsat 8 launch). **1.5. Removal of pre-fire forest attribute layers** Pre-fire forest attribute layers (e.g., canopy density, biomass, species composition) are no longer included in this version of CanLaBS. These attributes are now provided through the Spatialized Canadian National Forest Inventory (SCANFI v2; Guindon et al., 2026 ), which offers a more comprehensive, internally consistent, and regularly updated source of pre-disturbance forest information. Users are encouraged to combine CanLaBS with SCANFI v2 (Guindon et al., 2026) for their analyses. Users should use forest attributes from 2 years before the fire to avoid over-smoothed data that artificially underestimate pre-fire forest vegetation when pre-fire year Landsat data are unavailable. The fire start dates can be accessed via NBAC (https://cwfis.cfs.nrcan.gc.ca/datamart). **2. Use limitations** 2.1. This database is not designed to study a single fire or a limited number of fires but rather to study large areas with several fires. No radiometric correction or change was made per fire such as the offset method, or a mean, or median approach for pixels of the same year (see cjfr-2020-0353supplb at https://doi.org/10.1139/cjfr-2020-0353). Even if surface reflectance images were used, there may be radiometric differences within the same fire due to the use of different Landsat scenes. Differences in atmospheric correction between adjacent scenes may therefore be perceptible. The primary reason for not applying additional corrections in these cases is the insufficient number of pixels available per fire during July and August, particularly in certain regions and specific time periods.To achieve a spatially and temporally consistent database, a uniform processing approach was applied to all pixels. These points are discussed in the article and in the supplementary material (see cjfr-2020-0353supplb at https://doi.org/10.1139/cjfr-2020-0353). 2.2. Burnt areas that have undergone salvage logging were detected using a classification approach. This is not an exhaustive mapping of all areas that were salvage logged beyond one year after the fire, the goal was to eliminate these areas from the analyses, as the post-fire values (NBRpost) would be biased by the absence of trees and by the presence of soil disturbed by scarification. 2.3. Fires occurring in forests heavily affected by the mountain pine beetle (Dendroctonus ponderosae), spruce budworm (Choristoneura fumiferana), or other defoliators should ideally be excluded from analyses, as pre-fire NBR values are inherently low, potentially biasing dNBR-based assessments. CanLaD (Perbet et al., 2025) now provides identification of these affected areas (available at https://doi.org/10.23687/902801fd-4d9d-4df4-9e95-319e429545cc). 2.4. The 1985 and 2024 fires represent the beginning and end years of the time series, it is possible that some fires are incomplete for these years, and perhaps to a lesser extent for the 1986 and 2023 fires. **3. Summary** Overall, this update improves the precision and temporal coverage of the CanLaBS data product by leveraging Landsat Collection 2 with updated national fire perimeter polygons and a refined salvage detection method. These changes enhance the suitability of the dataset for national-scale analyses of fire effects, post-fire management, and long-term disturbance dynamics in Canadian forests. **4. Layers description** There are 3 layers: - CanLaBS_1985_2024_v20260121.tif - dNBR values for all burnt pixels according to NBAC - CanLaBS_salvageMask_1985_2024_v20260121.tif - Binary layer where '1' identifies pixels where salvage logging occurred - NBAC_MRB_1972to2024_reproj.tif - NBAC fire year **5. Data download** The data can be downloaded from the FTP server (ftp.maps.canada.ca/pub/nrcan_rncan/Forest-fires_Incendie-de-foret/CanLaBS_v2-Burned_Severity-Severite_des_feux), referenced in the “Data and Resources” section, using a browser download manager, such as DownThemAll, or an external client such as FileZilla. **6. Dataset citation** - Guindon L., Correia D., Perbet P. 2026. Canada Landsat Burned Severity (CanLaBS v2): a Canada-wide Landsat-based 30-m resolution product of burned severity since 1985. https:/doi.org/10.23687/2af751e7-79f9-4da8-9b45-14688818dca3 **7. References** - Earth Resources Observation and Science (EROS) Center. 2020a. Landsat 4–5 Thematic Mapper Level-2, Collection 2. Dataset. U.S. Geological Survey. https://doi.org/10.5066/P9IAXOVV - Earth Resources Observation and Science (EROS) Center. 2020b. Landsat 7 Enhanced Thematic Mapper Plus Level-2, Collection 2. Dataset. U.S. Geological Survey. https://doi.org/10.5066/P9C7I13B - Earth Resources Observation and Science (EROS) Center. 2020c. Landsat 8–9 Operational Land Imager / Thermal Infrared Sensor Level-2, Collection 2. Dataset. U.S. Geological Survey. https://doi.org/10.5066/P9OGBGM6 - Guindon, L., S. Gauthier, F. Manka, M. A. Parisien, E. Whitman, P. Bernier, A. Beaudoin, P. Villemaire, and R. Skakun. 2021. “Trends in Wildfire Burn Severity across Canada, 1985 to 2015.” Canadian Journal of Forest Research 51 (9): 1230–1244. https://doi.org/10.1139/cjfr-2020-0353 - Guindon, L., P. Villemaire, D. L. P. Correia, F. Manka, S. Lacarte, and B. Smiley. 2023. SCANFI: Spatialized CAnadian National Forest Inventory Data Product. Natural Resources Canada, Canadian Forest Service, Laurentian Forestry Centre, Quebec, Canada. https://doi.org/10.23687/18e6a919-53fd-41ce-b4e2-44a9707c52dc - Guindon, L., D. L. P. Correia, F. Manka, and B. Smiley. 2026. SCANFI v2: Spatialized Canadian National Forest Inventory Data Product. Quebec, Canada: Natural Resources Canada, Canadian Forest Service, Laurentian Forestry Centre. https://doi.org/10.23687/07653869-f303-46c2-a04e-9ab479b73cbf - Perbet, P., L. Guindon, D. L. P. Correia, et al. 2025. “Historical Insect Disturbance Maps from 1985 Onwards for Canadian Forests Derived Using Earth Observation Data.” Scientific Data 12: 2012. https://doi.org/10.1038/s41597-025-06269-x - Perbet, P., L. Guindon, D. L. P. Correia, P. Villemaire, O. Reisi Gahrouei, and R. St-Amant. Canada Landsat Disturbance with Pest (CanLaD): A Canada-Wide Landsat-Based 30-m Resolution Product of Fire, Harvest and Pest Outbreak Detection and Attribution since 1987. https://doi.org/10.23687/902801fd-4d9d-4df4-9e95-319e429545cc - Skakun, R., G. Castilla, J. Metsaranta, E. Whitman, S. Rodrigue, J. Little, K. Groenewegen, and M. Coyle. 2022. “Extending the National Burned Area Composite Time Series of Wildfires in Canada.” Remote Sensing 14 (13): 3050.

This data publication includes three map datasets: 1. Detailed annual disturbance maps – Forty raster layers mapping Canadian forest disturbance types from 1985 to 2024 at a 30m resolution. 2. Latest disturbance type and latest disturbance year maps – A simplified version for enhanced usability, consisting of three rasters that represent the most recent disturbance type and year (starting and ending). 3. Latest Landsat time series Julian day map – A raster capturing the latest Julian day for each pixel in the time series, where the Julian day represents the number of days elapsed since January 1, 1970. The forest disturbance types include wildfire, harvesting, pest outbreaks, windthrow, and new water bodies. The method is based on the summer composite Landsat time series (Guindon et al 2024). Disturbance breaks in the time series are first detected using the LandTrendr approach (Kennedy et al., 2010). Next, a one-dimensional convolutional neural network model (TempCNN; Pelletier et al., 2019) is applied with a 10-year window to classify disturbance types. The resulting maps achieve an overall accuracy of 81%. For aggregated pest/no-pest classes, the overall accuracy is 88.0% ±1.2%, with a commission error of 25.4% ±5.7% and an omission error of 63.2% ±4.3%. ## Data use constraints: 1. For pest-related disturbances, the proposed defoliation severity classes do not directly correspond to annual aerial survey classifications. Instead, they represent the intensity of cumulative spectral change at the end of the epidemic or, for ongoing outbreaks, the most recent observed year. Contrary to aerial survey classifications, only moderate to severe cumulative defoliation levels are detected, representing a good compromise between omission error and commission. 2. The models are aimed at insect pests that primarily affect conifers. However, they may also capture severe defoliation in mixed or deciduous forests. 3. The analysis is based on a 10-year time window to adequately capture the effects of progressive defoliation. Therefore, to properly detect a pest causing this type of defoliation, such as the spruce budworm, historical data only becomes truly relevant starting around 1995. For insects with faster defoliation, the 1990s might be considered a good starting point. 4. The wood harvesting class refers to the removal of trees, regardless of the underlying intention. It primarily includes areas intended to remain forested or to be reforested but may also encompass certain sectors converted to other uses, such as road construction, mining, or various infrastructure projects. 6. The windthrow class is effective at detecting large-scale events but has a high false detection rate, particularly along the edges of harvested areas, where mixed pixels create spectral similarities with windthrow. Among all disturbance classes, windthrow has the highest error rate. 7. The new water body class was not formally validated in this study, though visual assessments were conducted. 8. Since the summer composite considers only July and August imagery, disturbances occurring in the fall are detected the following year. For example, a wildfire that occurred in August 2023 might only become visible in the 2024 composite if the 2023 composite used images from early July, before the disturbance occurred. Additionally, cloud or shadow masking can create gaps in the time series, causing some disturbance events to appear delayed by one or two years. Users can use the national fire database (NBAC Canadian Wildland Fire Information System) for validation and year adjustments of wildfires. Moreover, the last Julian days raster could be used to better interpret the predicted year of disturbance. 9. The most recent years in the time series may have a higher commission error. These errors will be addressed in future annual updates. 10. The minimum detectable disturbance size is 1.08 hectares (12 pixels), which may limit the detection of linear disturbances such as roads. ## More details will be available in the future scientific publication: Perbet, P., ## Please cite this dataset as: Perbet, P., Guindon, L., Correia D.L.P., P. Villemaire, O., Reisi Gahrouei R. St-Amant, Canada Landsat Disturbance with pest (CanLaD): a Canada-wide Landsat-based 30-m resolution product of fire, harvest and pest outbreak detection and attribution since 1987. https://doi.org/10.23687/902801fd-4d9d-4df4-9e95-319e429545cc ## Cited references: Guindon, Luc, Francis Manka, David L.P. Correia, Philippe Villemaire, Byron Smiley, Pierre Bernier, Sylvie Gauthier, Andre Beaudoin, Jonathan Boucher, et Yan Boulanger. 2024. « A New Approach for Spatializing the CAnadian National Forest Inventory (SCANFI) Using Landsat Dense Time Series ». Canadian Journal of Forest Research, février, cjfr-2023-0118. https://doi.org/10.1139/cjfr-2023-0118. Kennedy, Robert E., Zhiqiang Yang, et Warren B. Cohen. 2010. « Detecting trends in forest disturbance and recovery using yearly Landsat time series: 1. LandTrendr — Temporal segmentation algorithms ». Remote Sensing of Environment 114 (12): 2897 2910. https://doi.org/10.1016/j.rse.2010.07.008. Pelletier, Charlotte, Geoffrey I. Webb, et François Petitjean. 2019. « Temporal Convolutional Neural Network for the Classification of Satellite Image Time Series ». Remote Sensing 11 (5): 523. https://doi.org/10.3390/rs11050523.

The fire regime describes the patterns of fire seasonality, frequency, size, spatial continuity, intensity, type (e.g., crown or surface fire) and severity in a particular area or ecosystem. Annual area burned is the average surface area burned annually in Canada by large fires (greater than 200 hectares (ha)). Changes in annual area burned were estimated using Homogeneous Fire Regime (HFR) zones. These zones represent areas where the fire regime is similar over a broad spatial scale (Boulanger et al. 2014). Such zonation is useful in identifying areas with unusual fire regimes that would have been overlooked if fires had been aggregated according to administrative and/or ecological classifications. Fire data comes from the Canadian National Fire Database covering 1959–1999 (for HFR zones building) and 1959-1995 (for model building). Multivariate Adaptive Regression Splines (MARS) modeling was used to relate monthly fire regime attributes with monthly climatic/fire-weather in each HFR zone. Future climatic data were simulated using the Canadian Earth System Model version 2 (CanESM2) and downscaled at a 10 Km resolution using ANUSPLIN for two different Representative Concentration Pathways (RCP). RCPs are different greenhouse gas concentration trajectories adopted by the Intergovernmental Panel on Climate Change (IPCC) for its fifth Assessment Report. RCP 2.6 (referred to as rapid emissions reductions) assumes that greenhouse gas concentrations peak between 2010-2020, with emissions declining thereafter. In the RCP 8.5 scenario (referred to as continued emissions increases) greenhouse gas concentrations continue to rise throughout the 21st century. Multiple layers are provided. First, the annual area burned by large fires (>200 ha) is shown across Canada for a reference period (1981-2010). Projected annual area burned layers are available for the short- (2011-2040), medium- (2041-2070), and long-term (2071-2100) under the RCP 8.5 (continued emissions increases) and, for the long-term (2071-2100), under RCP 2.6 (rapid emissions reductions). Reference: Boulanger, Y., Gauthier, S., et al. 2014. A refinement of models projecting future Canadian fire regimes using homogeneous fire regime zones. Canadian Journal of Forest Research 44, 365–376.

The fire regime describes the patterns of fire seasonality, frequency, size, spatial continuity, intensity, type (e.g., crown or surface fire) and severity in a particular area or ecosystem. Annual area burned is the average surface area burned annually in Canada by large fires (greater than 200 hectares (ha)). Changes in annual area burned were estimated using Homogeneous Fire Regime (HFR) zones. These zones represent areas where the fire regime is similar over a broad spatial scale (Boulanger et al. 2014). Such zonation is useful in identifying areas with unusual fire regimes that would have been overlooked if fires had been aggregated according to administrative and/or ecological classifications. Fire data comes from the Canadian National Fire Database covering 1959–1999 (for HFR zones building) and 1959-1995 (for model building). Multivariate Adaptive Regression Splines (MARS) modeling was used to relate monthly fire regime attributes with monthly climatic/fire-weather in each HFR zone. Future climatic data were simulated using the Canadian Earth System Model version 2 (CanESM2) and downscaled at a 10 Km resolution using ANUSPLIN for two different Representative Concentration Pathways (RCP). RCPs are different greenhouse gas concentration trajectories adopted by the Intergovernmental Panel on Climate Change (IPCC) for its fifth Assessment Report. RCP 2.6 (referred to as rapid emissions reductions) assumes that greenhouse gas concentrations peak between 2010-2020, with emissions declining thereafter. In the RCP 8.5 scenario (referred to as continued emissions increases) greenhouse gas concentrations continue to rise throughout the 21st century. Provided layer: projected annual area burned by large fires (>200 ha) across Canada for the medium-term (2041-2070) under the RCP 8.5 (continued emissions increases). Reference: Boulanger, Y., Gauthier, S., et al. 2014. A refinement of models projecting future Canadian fire regimes using homogeneous fire regime zones. Canadian Journal of Forest Research 44, 365–376.

The fire regime describes the patterns of fire seasonality, frequency, size, spatial continuity, intensity, type (e.g., crown or surface fire) and severity in a particular area or ecosystem. The number of large fires refers to the annual number of fires greater than 200 hectares (ha) that occur per units of 100,000 ha. It was calculated per Homogeneous Fire Regime (HFR) zones. These HFR zones represent areas where the fire regime is similar over a broad spatial scale (Boulanger et al. 2014). Such zonation is useful in identifying areas with unusual fire regimes that would have been overlooked if fires had been aggregated according to administrative and/or ecological classifications. Fire data comes from the Canadian National Fire Database covering 1959–1999 (for HFR zones building) and 1959-1995 (for model building). Multivariate Adaptive Regression Splines (MARS) modeling was used to relate monthly fire regime attributes with monthly climatic/fire-weather in each HFR zone. Future climatic data were simulated using the Canadian Earth System Model version 2 (CanESM2) and downscaled at a 10 Km resolution using ANUSPLIN for two different Representative Concentration Pathways (RCP). RCPs are different greenhouse gas concentration trajectories adopted by the Intergovernmental Panel on Climate Change (IPCC) for its fifth Assessment Report. RCP 2.6 (referred to as rapid emissions reductions) assumes that greenhouse gas concentrations peak between 2010-2020, with emissions declining thereafter. In the RCP 8.5 scenario (referred to as continued emissions increases) greenhouse gas concentrations continue to rise throughout the 21st century. Provided layer: the number of large fires (>200 ha) across Canada for a reference period (1981-2010). Reference: Boulanger, Y., Gauthier, S., et al. 2014. A refinement of models projecting future Canadian fire regimes using homogeneous fire regime zones. Canadian Journal of Forest Research 44, 365–376.

The fire regime describes the patterns of fire seasonality, frequency, size, spatial continuity, intensity, type (e.g., crown or surface fire) and severity in a particular area or ecosystem. The number of large fires refers to the annual number of fires greater than 200 hectares (ha) that occur per units of 100,000 ha. It was calculated per Homogeneous Fire Regime (HFR) zones. These HFR zones represent areas where the fire regime is similar over a broad spatial scale (Boulanger et al. 2014). Such zonation is useful in identifying areas with unusual fire regimes that would have been overlooked if fires had been aggregated according to administrative and/or ecological classifications. Fire data comes from the Canadian National Fire Database covering 1959–1999 (for HFR zones building) and 1959-1995 (for model building). Multivariate Adaptive Regression Splines (MARS) modeling was used to relate monthly fire regime attributes with monthly climatic/fire-weather in each HFR zone. Future climatic data were simulated using the Canadian Earth System Model version 2 (CanESM2) and downscaled at a 10 Km resolution using ANUSPLIN for two different Representative Concentration Pathways (RCP). RCPs are different greenhouse gas concentration trajectories adopted by the Intergovernmental Panel on Climate Change (IPCC) for its fifth Assessment Report. RCP 2.6 (referred to as rapid emissions reductions) assumes that greenhouse gas concentrations peak between 2010-2020, with emissions declining thereafter. In the RCP 8.5 scenario (referred to as continued emissions increases) greenhouse gas concentrations continue to rise throughout the 21st century. Provided layer: projected number of large fires (>200 ha) across Canada for the long-term (2071-2100) under the RCP 2.6 (rapid emissions reductions). Reference: Boulanger, Y., Gauthier, S., et al. 2014. A refinement of models projecting future Canadian fire regimes using homogeneous fire regime zones. Canadian Journal of Forest Research 44, 365–376.

The Regional Air Quality Deterministic Prediction System FireWork (RAQDPS-FW) carries out physics and chemistry calculations, including emissions from active wildfires, to arrive at deterministic predictions of chemical species concentration of interest to air quality, such as fine particulate matter PM2.5 (2.5 micrometers in diameter or less). Geographical coverage is Canada and the United States. Data is available at a horizontal resolution of 10 km. While the system encompasses more than 80 vertical levels, data is available only for the surface level. The products are presented as historical, annual or monthly, averages which highlight long-term trends in cumulative effects on the environment.