| 10 Dec 2025
A Global 30 m Landsat-based Dataset of Forest Fire Patches (GlobMap FFP v1.0) from 1984 to 2022
Abstract. Forest fires exert profound ecological impacts globally. Characterizing their long-term effects and evolving regimes requires consistent, high-resolution fire records over extended periods. The Landsat archive provides a unique foundation for such efforts, offering fine spatial detail with globally coherent, multi-decadal observations. Yet, it remains challenging to generate a globally consistent, Landsat-based fire product with event-level characterization. Here we present a 30 m global forest fire patch dataset spanning 1984–2022, developed from the full Landsat archive to ensure comprehensive fire characterization. We first condensed multi-temporal burned signals from Landsat archive on Google Earth Engine (GEE) using a pixel-based image compositing approach. This approach also reduces noise from clouds and shadows while ensuring high computational efficiency using GEE. We then mapped burned area using artificial neural network modeling across global forests. Finally, we delineated individual fire patches through spatial–temporal clustering and extracted their key attributes. Across global forests, we identified a total of 11.97 million individual fire patches burning 7.3 Mha yr−1over 1984–2022. Validation indicated omission errors ranging from 12.2 % to 36.8 % and commission errors ranging from 6.4 % to 23.2 % across diverse forest types. Intercomparison with existing Landsat products revealed strong agreement in annual burned area estimates and fire patch detection, with discrepancies mainly arising from within-fire delineation and small fire detection. This dataset offers a valuable resource for quantifying fire impacts and advancing the understanding of contemporary and future fire regimes in global forests.
This manuscript presents GlobMap FFP v1.0, a global 30 m dataset of forest burned area patches derived from the Landsat archive for the period 1984–2022. Although the technical effort required to process the full Landsat archive at global scale is substantial, the study is built on a fundamentally flawed premise: Landsat’s temporal resolution is insufficient for reliable global burned area mapping. Burned area is a phenomenon characterized by short-lived spectral signals, often detectable for only a few weeks to a few months, and strongly dependent on observation timing and cloud-free conditions. A sensor with a nominal 16-day revisit cycle, further reduced by cloud cover and data gaps, cannot consistently observe this phenomenon at global scale, particularly in tropical and temperate forests.
The manuscript implicitly acknowledges this limitation by adopting multi-year temporal compositing, typically using five-year windows (and even longer periods prior to 2000). However, this strategy does not solve the underlying problem. Aggregating observations over five-year intervals inevitably suppresses short-lived burn signals and biases detection toward high-severity or slowly recovering fires. As a result, large fractions of real burned area are missed, especially in regions with frequent low-intensity fires or rapid vegetation recovery. An additional unresolved issue concerns areas that burn multiple times within the same compositing window. If I understand the methodology correctly, how are repeated fire occurrences affecting the same pixel over a five-year period handled, and how is fire recurrence represented when only a single observation per pixel is retained in the composite?
A critical consequence of this approach is evident in the reported burned area estimates. The dataset reports mean global forest burned area values of approximately 7.3 Mha yr⁻¹ for the period 2001–2021, whereas established global burned area products report values close to 19–20 Mha yr⁻¹ over the same period. This represents an underestimation by a factor of roughly three.
This strong underestimation is particularly pronounced in tropical regions, where the manuscript itself reports the largest divergences relative to existing products. These are precisely the regions where burned signals are short-lived, cloud cover is persistent, and Landsat’s sparse clear-sky observations are least capable of capturing fire effects. Therefore, I find it very difficult to reconcile the large disparity in burned area between the GlobMap product and MODIS with the validation metrics reported in the manuscript. According to Table 2, omission errors (24%) and commission errors (13%) are relatively low, the Dice coefficient is high (0.82), and the reported relative bias is relatively small (−11%). These values would normally indicate a product with only moderate underestimation. However, the intercomparison shows that GlobMap detects substantially less burned area (Fig. 7; MODIS ≈ 19.6 Mha yr⁻¹ versus GlobMap ≈ 7.3 Mha yr⁻¹). If such a large and systematic discrepancy truly exists, it is difficult to explain how it could coexist with a low relative bias and only moderate omission errors. This inconsistency is not resolved in the manuscript and fundamentally undermines the credibility of the validation results.
The validation framework itself further limits the interpretability of the reported accuracy metrics. Although the authors state that they follow the spatial sampling framework of the Burned Area Reference Database (BARD) by selecting Landsat TSAs, they do not use the BARD reference perimeters. Instead, reference burned area within each TSA is generated independently using the same algorithmic approach employed for training sample generation and product development. As a result, the validation does not rely on independent reference data, but rather evaluates internal methodological consistency.
Moreover, while BARD TSAs are part of a global, stratified sampling design explicitly constructed to support statistically rigorous accuracy assessment and uncertainty estimation, this sampling design is not adopted in the present study. The authors do not implement a probability-based sampling scheme adapted to their analysis, nor do they report uncertainty measures (e.g. confidence intervals or standard errors) for the accuracy metrics. Consequently, the reported omission and commission errors, Dice coefficients, and relative bias cannot be interpreted as statistically robust estimates of real-world burned area detection performance.
In addition, the validation methodology lacks essential temporal detail. The manuscript does not clearly specify which years or portions of the 1984–2022 period are actually covered by the validation, nor whether the reported accuracy metrics are representative of the entire time series. It remains unclear whether the validation is dominated by periods with higher observation density (e.g. the Landsat 7 and Landsat 8 eras) or whether earlier periods, characterized by sparser data availability and longer compositing windows, are adequately represented. Furthermore, the additional filtering of validation imagery to scenes with less than 40% cloud cover further reduces the number of usable observations, particularly in cloud-prone regions, compounding uncertainties regarding the representativeness and robustness of the reported metrics.
An additional conceptual limitation concerns the interpretation of the mapped units as 'individual' fire events. In the proposed dataset, burned pixels are aggregated into patches based on spatial proximity and assignment to a single burned year, without explicit information on ignition timing, fire duration, or intra-annual separation. Under this framework, independent fires occurring at different moments within the same year may be merged into a single patch, while the same fire spreading over extended periods may be inconsistently represented depending on observation availability. As a result, the mapped patches cannot be unambiguously interpreted as fire events. This has direct implications for the analysis of fire size distributions, fire frequency, and fire regime characteristics, and further limits the comparability of the dataset with products that explicitly track fire events using finer temporal information. The manuscript does not sufficiently clarify these limitations or their consequences for downstream analyses.
Overall, while Landsat-based burned area mapping can be highly effective at regional scales under appropriate conditions, this manuscript does not demonstrate that such approaches can be straightforwardly generalized to a global product without substantial loss of information. The strong underestimation of burned area, the internal inconsistency between area estimates and validation metrics, the lack of independent validation, and the inability to represent repeated burning indicate that the proposed dataset does not provide a reliable or improved representation of global forest burned area. In this context, it remains unclear what scientific or practical value a new dataset based on higher spatial resolution sensors offers if it does not demonstrably improve the representation of burned area relative to existing products.