the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 12 May 2026
A Moderate-Resolution, Long-Term Global Radar-Based Forest Above-Ground Biomass Dataset from 1993 to 2020
Abstract. Understanding global carbon dynamics and budgets under climate change, land-use shifts, and increasing disturbances remains challenging due to the limitations of existing coarse spatial resolution and short-term or discontinuous biomass datasets. In this study, we generated a new global annual forest above-ground biomass dataset at 8.9 km spatial resolution from 1993 to 2020. This dataset is derived from satellite radar backscatter data and integrates background climate constraints to accounts for regional differences and improve the accuracy of global above-ground biomass mapping. Our dataset estimates an average global forest above-ground biomass carbon stock of 191 ± 2.5 PgC, aligning with other global estimates. We observed an increase in global forest above-ground biomass carbon stocks from 1993 to 2020 at a rate of 0.29 PgC yr−1. Tropical Africa, temperate and boreal forests are the primary contributors to global forest above-ground biomass carbon stock gains from 1993 to 2020. In contrast, gross above-ground biomass carbon losses are predominantly observed in tropical America and Asia forests, particularly since 2000. This long-term, temporally continuous, and moderate-resolution dataset provides a new benchmark for quantifying biomass carbon dynamics and integrating these processes into Earth System Models. The above-ground forest biomass dataset is openly accessible, alongside this manuscript.
- Preprint
(4209 KB) - Metadata XML
-
Supplement
(3471 KB) - BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on essd-2026-246', Anonymous Referee #1, 09 Jun 2026
-
RC2: 'Comment on essd-2026-246', Anonymous Referee #2, 31 Aug 2026
The comment was uploaded in the form of a supplement: https://essd.copernicus.org/preprints/essd-2026-246/essd-2026-246-RC2-supplement.pdf
-
EC1: 'Comment on essd-2026-246', Nophea Sasaki, 09 Sep 2026
The manuscript presents a potentially valuable global annual forest above-ground biomass (AGB) record for 1993-2020. A record of this duration could become a useful community resource. However, I agree with both referees that substantial revision and additional validation are required before the product can be considered reliable for interpreting interannual-to-decadal biomass change. The map-level agreement reported for biomass stocks does not, by itself, establish the fidelity of the temporal signal, which is the main novelty of the dataset.
My main editorial requirements, complementing the detailed referee reports, are as follows.
1. Evaluate spatial and temporal generalization separately: The selected model obtains most of its map-level predictive skill after the addition of static background climate variables, whereas all year-to-year variability must come from the radar record. The current performance statistics therefore mix spatial stock prediction with temporal-change prediction. Moreover, the public training code constructs cross-validation groups by combining year with 0.5-degree spatial blocks. This permits observations from the same spatial block in different years to occur in different folds; the 2020 holdout is temporally, but not spatially, independent. Please revise or clarify the evaluation so that spatial blocks are held out across all years, and add temporal and sensor-era holdouts designed specifically to test annual anomalies and trends. At minimum, report performance for anomalies and/or first differences, by biome and biomass range, in addition to performance for biomass levels.
2. Demonstrate temporal homogeneity and isolate the biomass signal: Please provide explicit diagnostics for all ERS, QSCAT, and ASCAT overlap and transition periods, not only the visible 1997 anomaly. These should include breakpoint or change-point tests, estimated offsets, any corrections applied, and a clear quality flag or exclusion rule for affected years. The authors should also show that the reconstructed anomalies and trends remain after accounting for soil moisture and vegetation-water effects. Useful tests could include comparisons in demonstrably stable forest, areas with independently documented disturbances or recovery, and anomaly/partial-correlation analyses against radar and soil-moisture series. The central criterion is whether the radar-derived temporal signal represents biomass change rather than hydrological variability or sensor discontinuities.
3. Clarify the quantities called gross gain, gross loss, and net sink: As presently described, gains and losses appear to be aggregated after pixels are classified by their full-period positive or negative trend. If so, these quantities are a decomposition by a posteriori trend class, not independent estimates of annual gross biomass gains and losses. Please define the calculation precisely, discuss the selection implied by this classification, and use terminology that cannot be confused with gross carbon fluxes. The manuscript should also reconcile the reported global AGC rates of 0.29 and 0.44 PgC yr-1 by stating the estimator, period, and spatial mask used for each. Causal statements attributing regional changes to CO2 fertilization, management, drought, fire, or land-use change should be framed as hypotheses unless attribution is directly tested.
4. Correct the treatment of resolution and uncertainty: Please distinguish the 8.9 km grid spacing from the effective spatial resolution supported by the native scatterometer footprints and the merged radar product, and revise claims about detecting small disturbances accordingly. Equation (1), Figure 5, the error bars, and the quoted +/-2.5 PgC must be made mutually consistent: the equation appears to describe variability in fold-level mean error, whereas Figure 5 is spatially varying. The revised paper and data release should state exactly what each uncertainty quantity represents and should provide per-cell uncertainty and quality information. Uncertainty or sensitivity in stock and trend estimates should address, as far as possible, training-product bias, biomass-range compression evident against inventories, sensor harmonization, forest masks and exclusions, carbon fraction, and the root-to-shoot conversion. Claims that the product is a "benchmark" should be tempered unless these uncertainties are adequately characterized.
5. Make the archived product and code independently usable and reproducible: The current Zenodo record provides a single compressed NetCDF with only a brief landing-page description, while the GitHub repository contains six workflow scripts and a short overview; the scripts use project-specific paths and no reproducible software environment is supplied. Before acceptance, please provide a data user guide and complete NetCDF metadata, including variable definitions, units, coordinates/CRS, cell area, forest-mask definition, fill values, valid ranges, uncertainty fields, and sensor/year quality flags. The code release should include an open-source license, pinned environment or dependency file, configurable paths, exact commands and processing order, input dataset versions, model settings and random seeds, and an archived release tied to the manuscript DOI or an immutable commit. If model files or intermediate inputs are required to reproduce the published NetCDF, they should be archived or their generation documented unambiguously.
Please also complete a careful consistency and language revision, including the discrepancies already identified by the referees (for example, test R2 = 0.833 versus 0.883; six versus seven biome groups; C-band-only language despite the merged C- and Ku-band record; and terminology such as "continuous," "dynamic," "pixel," and "grid cell").
I would be willing to reconsider a substantially revised manuscript that addresses both referee reports and the points above. The decisive issue is not whether the model reproduces broad spatial biomass gradients, but whether the authors can establish a homogeneous, uncertainty-characterized, and independently usable annual record of biomass change.
Citation: https://doi.org/10.5194/essd-2026-246-EC1
Data sets
A Moderate-Resolution, Long-Term Global Radar-Based Forest Above-Ground Biomass Dataset from 1993 to 2020 Liu Guohua, Ciais Philippe, Tao Shengli, Yang Hui, Xiao Chenwei, and Bastos Ana https://doi.org/10.5281/zenodo.19259660
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 435 | 180 | 66 | 681 | 75 | 61 | 85 |
- HTML: 435
- PDF: 180
- XML: 66
- Total: 681
- Supplement: 75
- BibTeX: 61
- EndNote: 85
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The authors present a new global forest above-ground biomass dataset spanning 1993–2020 at 8.9 km resolution, produced by training a random forest model on ESA-CCI biomass maps using radar backscatter and climate covariates as predictors. Compared to existing long-term products, which are VOD-based at ~25 km, the dataset offers a finer spatial resolution while retaining a near-30-year continuous annual record. This is achieved by exploiting the recently published radar backscatter record of Tao et al. (2023). The product is potentially a useful community resource, and the manuscript includes comparisons against other biomass products and field inventories. However, two important issues should be addressed before publication, which I outline below.
1. The product's central contribution is temporal — interannual to decadal biomass dynamics — but the model structure raises a concern about where that temporal signal originates. Two of the three predictors (MAT_bg, MAP_bg) are static background climate and cannot, by construction, generate any interannual variation; the authors further show that adding dynamic climate does not improve performance (Table 1). The temporal dynamics in the product therefore rest entirely on the radar backscatter input. Radar backscatter is, by its nature, sensitive to soil moisture and vegetation water content as well as to biomass; the authors' wet-season index reduces but does not remove this, and even the selected index retains a non-trivial correlation with soil moisture (Figure S3). Because the product's interannual signal comes only from radar, any residual moisture variability will project directly onto the reported biomass trends. Combined with the low standalone radar skill (test R² = 0.136), this raises the concern that part of the temporal signal may reflect moisture dynamics rather than biomass change. To be clear, I am not claiming radar is uninformative — with correlated predictors the variance explained is shared between radar and climate, and the large jump to R² = 0.833 when static climate is added may reflect climate carrying much of the spatial structure. The concern is specific to the temporal signal that distinguishes this dataset. The authors should (i) report variable importance for the selected model, and (ii) demonstrate that the interannual and decadal trends are driven by genuine radar-sensed biomass change rather than residual moisture variability — for example by relating the temporal anomalies back to the radar signal and to soil moisture directly.
2. The reported net sink of 0.44 PgC yr⁻¹ is a small residual between much larger, oppositely-signed gross gain (~0.84 PgC yr⁻¹) and gross loss (~−0.39 to −0.74 PgC yr⁻¹) terms, which makes it sensitive to systematic biases in the product. The quoted uncertainty of ±2.5 PgC (Eq. 1) captures only the spread across cross-validation iterations and does not propagate other potentially important sources — calibration bias (the slope of 0.36 against inventories), offsets at the ERS/ASCAT/QSCAT sensor transitions, and the fixed conversion factors (0.49 carbon fraction; root-to-shoot map) applied globally. It would strengthen the paper if the authors acknowledged more explicitly that the reported uncertainty is a lower bound reflecting model variance only, discussed these additional sources qualitatively even if they cannot be fully quantified, and tempered the net-flux claims accordingly. The 1997 artefact (Figure 3B) is a useful illustration here: it shows that a sensor transition can produce a spurious global biomass swing, so a brief comment on whether the other transitions were checked for similar (smaller) offsets would help reassure readers about the homogeneity of the trend record.
Minors comments
C-band saturation in dense forest acknowledged (l. 338–341) but underplayed given tropics = ~66% of stock.
Test R² inconsistent: 0.833 (Table 1) vs 0.883 (l. 213).