Source-linked AI summary
30m resolution Global Annual Burned Area Mapping based on Landsat images and Google Earth Engine
Tengfei Long, Zhaoming Zhang, Guojin He, Weili Jiao, Chao Tang, Bingfang Wu, Xiaomei Zhang, Guizhou Wang, Ranyu Yin
TL;DR
Global burned area products have lacked the spatial detail of a global 30-meter product, and Landsat-based mapping is challenged by limited temporal resolution and rapid vegetation recovery. This study uses time-series Landsat imagery and Google Earth Engine to build an automated global annual mapping pipeline and release GABAM 2015. GABAM 2015 achieved 13.17% commission error and 30.13% omission error, while comparison with Fire cci showed similar spatial distribution and strong correlation.
Problem
Global burned area products lacked sufficient spatial detail, and no global Landsat-based burned area product had been reported.
Method
The study develops an automated global annual burned-area mapping pipeline using time-series Landsat imagery and Google Earth Engine.
Results
13.17% commission error and 30.13% omission error were reported for GABAM 2015, with similar spatial distribution and strong correlation against Fire cci.
Takeaways & Limitations
GABAM 2015 provides a novel 30-meter global annual burned-area map for fires occurring during 2015.
Takeaways & Limitations
In tropical validation sites, rapid vegetation recovery caused clear burned evidence to be missed by Landsat, contributing to omission error.
Abstract
from arXiv · showhide
Heretofore, global burned area (BA) products are only available at coarse spatial resolution, since most of the current global BA products are produced with the help of active fire detection or dense time-series change analysis, which requires very high temporal resolution. In this study, however, we focus on automated global burned area mapping approach based on Landsat images. By utilizing the huge catalog of satellite imagery as well as the high-performance computing capacity of Google Earth Engine, we proposed an automated pipeline for generating 30-meter resolution global-scale annual burned area map from time-series of Landsat images, and a novel 30-meter resolution global annual burned area map of 2015 (GABAM 2015) is released. GABAM 2015 consists of spatial extent of fires that occurred during 2015 and not of fires that occurred in previous years. Cross-comparison with recent Fire_cci version 5.0 BA product found a similar spatial distribution and a strong correlation ($R^2=0.74$) between the burned areas from the two products, although differences were found in specific land cover categories (particularly in agriculture land). Preliminary global validation showed the commission and omission error of GABAM 2015 are 13.17% and 30.13%, respectively.
1. Introduction
Global burned area products have lacked sufficient spatial detail, while Landsat offers suitable spatial and temporal resolution for burned-area monitoring. The study addresses the absence of a global Landsat-based product by using Google Earth Engine to generate an automated 30-meter annual map for 2015.
- Accurate burned-area data support analyses of fire occurrence, fire impacts, and carbon emissions from biomass burning.
- Landsat provides a long burned-area record from the 1970s onward with spatial and temporal resolutions suited to science and management applications.
- Existing global burned area products are generally coarse, while 30-meter products have been limited to specific regions.
- Traditional global burned-area methods rely on high-temporal-resolution data, whereas Landsat’s limited temporal resolution complicates direct transfer of those approaches.
- Dense Landsat time series and hundreds of thousands of scenes create substantial processing demands for global annual mapping.
- The study uses Google Earth Engine’s imagery catalog and global-scale computing capabilities to produce an automated 30-meter global annual burned-area map and release GABAM 2015.
2. Methodology
The methodology combines stratified sampling across land-cover and burned-area-density strata with Landsat-8 feature extraction and Google Earth Engine processing. It evaluates spectral features and uses burned-surface variability to inform global burned-area mapping.
- Sampling design: The sampling design stratifies sites by seven land-cover categories and five burned-area-density levels, producing 35 strata.Land-cover categories are intersected with GFED4 2015 burned-area-density intervals to represent different fire frequencies and biomes.
- Sampling design: 120 Landsat scenes were selected for training and 80 validation sites were independently sampled across burned-area-density levels.Validation sites were kept at least 200 km from training samples to reduce overlap with training scenes.
- Feature preparation: The classifier used 12,881 expert-labeled points, including 6,735 burned and 6,146 unburned samples, from Landsat-8 imagery.Burned samples covered different burn severities and biomass types, while unburned samples represented vegetation, built-up land, bare land, shadows, and lake borders.
- Feature preparation: Six Landsat-8 surface-reflectance bands and burn-sensitive spectral indices were extracted for classification.The bands included BLUE, GREEN, RED, NIR, SWIR1, and SWIR2; indices included NBR, NBR2, BAI, and MIRBI.
- Feature selection: Post-fire spectral variability depends on pre-fire vegetation and combustion severity, so no spectral index is optimal across all environments and fire regimes.This variability motivated using multiple common spectral indices rather than relying on a single index.
- Feature selection: Spectral indices were generally more important than surface reflectance, with NBR2, BAI, MIRBI, and SAVI ranked most important by the random forest.All 14 evaluated Landsat features were retained as sensitive features for global burned-area mapping.
3.1. Product description
GABAM 2015 is a global annual burned-area map generated from Landsat time series, providing approximately 30-meter spatial resolution and isolating fires occurring during 2015. Temporal filtering helped distinguish new burns from unrecovered scars in example regions.
- Product scope: GABAM 2015 maps the spatial extent of fires occurring during 2015 at approximately 30-meter resolution.The product uses Geographic projection and is distributed as global 10×10-degree tiles.
- Product scope: The map was produced from dense Landsat time series using an automated pipeline implemented with Google Earth Engine.The product is based on annually processed Landsat imagery and is available for download.
- Global visualization: Burned-area density is defined as the proportion of burned pixels within each 0.25°×0.25° grid.This aggregation is used to visualize the global distribution of burned area.
- Example mapping: In a Canadian example, temporal filters removed previously burned scars and retained the component associated with new 2015 burning.The 2014 and 2015 annual composites revealed that one scar component predated the target year.
- Example mapping: The region-growing example progresses from a false-color Landsat image to burned-probability, seed, and final burned-area maps.The candidate seeds are expanded until no additional pixels meet the aggregation condition.
3.2. Comparison with Fire cci product
GABAM 2015 was compared with the approximately 250-meter Fire_cci version 5.0 product after compositing both datasets into annual burned-area representations. The products showed similar global distributions and a strong overall relationship, but Fire_cci generally reported more burned area and agreement varied by land-cover category.
- Data preparation: Annual Fire_cci pixel products were created by labeling pixels burned when any monthly Julian Day layer contained a valid detection.The annual grid product was then computed as the burned-pixel proportion in each 0.25°×0.25° grid.
- Pixel comparison: Both annual products detected burned areas in the Landsat example, but GABAM 2015 delineated finer burned-area boundaries than Fire_cci.Fire_cci’s coarser input resolution can classify mixed burned and unburned pixels as burned.
- Global comparison: The two products showed similar global burned-area distributions across 0.25°×0.25° grids.The comparison used annual grid compositions for GABAM and Fire_cci.
- Regression analysis: GABAM 2015 generally underestimated burned scars relative to Fire_cci, with regression slopes below 1 and intercepts near 0.Many grids had higher burned proportions in Fire_cci, consistent with differences in sensor spatial resolution.
- Regression analysis: The strongest relationships occurred in coniferous forest (R2 = 0.82), rangeland (R2 = 0.75), and shrub cover, whereas agriculture had R2 = 0.31 and “Others” had R2 = 0.07.The agriculture result was attributed to uncertainty in both products, while “Others” contained few burned grids.
- Regression analysis: The global relationship was mainly determined by rangeland, including woody savannas, savannas, and grasslands.Most burned areas were located in rangeland according to the scatter distributions.
3.3. Validation
Validation combined satellite imagery, fire-perimeter datasets, manually interpreted samples, and supervised classification across 80 sites. GABAM 2015 achieved 13.17% commission error, 30.13% omission error, and 93.92% overall accuracy, with omission driven by missed or weakly expressed burned pixels.
- Validation design: Validation used 80 sites and combined Landsat imagery with MTBS, CBERS-4, and Gaofen-1 reference sources.Reference-site dimensions varied by data source, and Landsat remained the majority source.
- Reference construction: Reference burned-area perimeters were generated from image pairs by manually selecting burned and unburned samples, applying SVM classification, and integrating detections annually.Experts subsequently reviewed and edited the resulting perimeters visually.
- Validation results: 13.17% commission error, 30.13% omission error, and 93.92% overall accuracy were reported for GABAM 2015.These statistics were derived from global cross-tabulation between GABAM and reference classifications.
- Validation results: Higher omission error was associated with rapidly recovering tropical vegetation and pixels within burned areas that lacked strong burned appearance.Such pixels could be excluded by GABAM while remaining part of complete burned scars in the reference data.
- Validation results: A validation site using MTBS perimeters recorded 1.45% commission error and 67.97% omission error.The example illustrates substantial site-level variation in the error measures.
3.4. Discussion
The discussion identifies temporal sampling, clouds, vegetation diversity, and cropland characteristics as constraints on Landsat-based global burned-area detection. Despite generally smaller mapped areas than Fire_cci, GABAM’s finer perimeters may support burned-biome area statistics and biomass-burning emission simulations.
- Temporal limitations: Landsat generally revisits a location at intervals exceeding 10 days, making direct active-fire observation unlikely when cloud coverage is considered.The discussion reports an active-fire observation probability below 10% under these conditions.
- Spectral limitations: Global detection without active-fire evidence is difficult because vegetation, phenology, burned-like land covers, and within-scar spectral properties vary widely.Burned scars may contain char, scorched vegetation, grass, or green leaves when fire severity is low.
- Potential improvements: Region-specific algorithms using fire behavior, land cover, and climate knowledge were proposed as a way to improve high-resolution global mapping.The discussion notes that the existing two-category vegetation discrimination is insufficient to separate burning types.
- Comparison with Fire_cci: GABAM generally mapped less burned area than Fire_cci because weakly expressed pixels within burned areas were excluded.This can increase omission for patch connectivity and completeness while preserving more detailed burned-area perimeters.
- Potential utility: GABAM’s detailed burned-area perimeters may support biome-level burned-area statistics and improve biomass-burning carbon-emission simulations.This consequence is stated within the discussion of the product’s finer spatial detail.
- Cropland limitations: Cropland remains difficult to classify because harvested or ploughed surfaces resemble burned areas and cropland fires are often small and short-lived.The authors note that cropland masks could help remove potential confusions.
- Temporal limitations: Landsat’s revisit gaps and cloud contamination can omit or weaken active- and post-fire evidence, especially where vegetation recovers quickly.The resulting temporal gaps are associated with higher omission error in tropical regions.
- Validation limitations: Validation relying mainly on Landsat limits assessment of sensor-specific inaccuracies such as radiometric, spectral, geolocation, and mixed-pixel effects.The authors therefore expect further extensive validation by professional interpreters.
4. Conclusions
The study presents an automated Google Earth Engine pipeline that generates 30-meter global annual burned-area maps from Landsat imagery. GABAM 2015 showed similar spatial distribution and strong correlation with Fire_cci, while global validation reported commission and omission errors of 13.17% and 30.13%.
- The study proposes an automated pipeline for generating 30m resolution global-scale annual burned-area maps using Landsat images and Google Earth Engine.
- GABAM 2015 is a novel 30-m resolution global annual burned-area map derived from available Landsat-8 images.
- 13.17% commission error and 30.13% omission error were reported for GABAM 2015 according to global validation.
- GABAM and Fire_cci showed similar spatial distributions and a strong correlation between their burned-area estimates, particularly in coniferous forests.
- The automated pipeline makes it possible to efficiently generate GABAM from the large catalog of Landsat images.
Appendix A. Examples of validation sites
The appendix presents site-validation examples using GF-1, CBERS-4, Landsat-8, and MTBS reference data. The examples compare reference burned-area maps or perimeters with burned areas detected by the proposed method.
- The appendix summarizes site-validation examples and their locations, reference-data sources, and error information.
- GF-1 reference burned-area maps are compared with burned areas detected by the proposed method.
- CBERS-4 reference burned-area maps are compared with burned areas detected by the proposed method.
- Landsat-8 examples from Africa and Australia compare reference burned-area maps with burned areas detected by the proposed method.
- One example compares MTBS 2015 perimeters, Landsat-8-derived reference perimeters, and burned areas generated by the proposed method.