A multi-temporal dataset for mapping burned areas in the Brazilian Cerrado using time series of remote sensing imagery
Alisson Cleiton de Oliveira, Thales Sehn Körting
cs.LG · cs.AI · cs.CV
Quick verdict
Open 3-year WFI Cerrado burned-area dataset with thorough RF tests; solid data paper, but self-drawn labels and omission/commission swapped in abstract.
Summary
This paper presents a multi-temporal tabular dataset for burned-area mapping over Chapada dos Veadeiros National Park (Goiás, Brazil), built from WFI sensor imagery aboard CBERS-4A, CBERS-4, and AMAZONIA-1 for 2020–2022. From 382 downloaded scenes, 235 cloud- and swath-filtered images were processed (band stacking, DOS atmospheric correction, resampling to 64 m, AROSICS co-registration) and aggregated — four bands plus five spectral indices — into a 500 m grid of 38,957 cells, with forward-fill replication to annual 365-day series. Cells were manually labeled totally burned (TB), partially burned (PB), or non-burned (NB), and Random Forest models were evaluated across three labeling and seven validation strategies using 200 Monte Carlo simulations each. The best strategy — train on three classes, then relabel TB as burned area — yielded overall IoU of 0.98/0.98/0.95 (2020/2021/2022) but burned-area IoU of only 0.66/0.71/0.49, with PB the weakest class (commission errors 0.64–0.73). Comparison with MCD64A1 gave errors of the same order as that product's own global validation, though the abstract's omission/commission figures appear transposed relative to Table 11. Main limitations: author-drawn reference labels from the same imagery, random spatial train/test splits prone to autocorrelation leakage, a single-model MCD64A1 comparison, and only RGB-NIR bands actually used for classification.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- The WFI constellation (CBERS-4, CBERS-4A, AMAZONIA-1) can support burned-area mapping with Random Forest on BGR-NIR time series without inter-satellite spectral calibration.
- Training RF with TB/PB/NB as three separate classes and then relabeling only TB as burned area (5th validation approach) yields the best results (overall IoU 0.98/0.98/0.95; BA-class IoU 0.66/0.71/0.49 for 2020/2021/2022).
- Partially burned cells are the hardest class, with commission errors of 0.644–0.725 across years (Table 9).
- Compared with MCD64A1, the 5th-approach maps show BA errors of the same order as MCD64A1's own global validation (Table 11: CE 0.227/0.282/0.593, OE 0.461/0.432/0.46).
- The labeled dataset and RF script are openly available for further machine learning exploration (ScienceDB DOI 10.57760/sciencedb.14040).
Strengths
Unique, open multi-sensor dataset
First systematic burned-area use of the CBERS-4/4A/AMAZONIA-1 WFI constellation: 235 quality-filtered scenes over 2020–2022, aggregated to a 239×163 grid of 500 m cells with four bands and five indices, plus manual TB/PB/NB labels, openly released on ScienceDB (DOI 10.57760/sciencedb.14040) with the RF script.
Systematic evaluation of labeling strategies
Three classification × seven validation approaches (Tables 5–8), each with 200 Monte Carlo simulations, yields an actionable finding: three-class training followed by relabeling only TB as BA (5th approach) maximizes BA IoU in 2020/2021 (0.66/0.71 vs 0.48–0.69 for alternatives).
Honest per-class error analysis
Contextualized comparison with MCD64A1
Fully documented preprocessing chain
Weaknesses
Reference labels are circular and unvalidated
Abstract transposes headline error metrics
Optimistic validation design
Dataset under-explored in its own paper
Recurring numeric inconsistencies
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper introduces an openly available, multi-temporal tabular dataset built from WFI sensor imagery (CBERS-4A, CBERS-4, AMAZONIA-1; 55/64 m, 1–3 day revisit) over Chapada dos Veadeiros National Park plus a 10 km buffer, 2020–2022. Four bands and five spectral indices (BAI, EVI, GEMI, NDVI, NDWI) are aggregated to a 500 m grid; cells carry manual TB/PB/NB labels. The authors claim: (i) the first systematic burned-area use of this constellation; (ii) that Random Forest performs acceptably without inter-satellite spectral calibration; (iii) that training on three classes and relabeling only TB as burned area (the 5th validation approach) is optimal. All three are plausible but only partially supported by the evidence presented.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does this paper's dataset contain?
A tabular time series (2020–2022) over Chapada dos Veadeiros National Park, Brazil: mean values of blue, green, red and NIR bands plus the BAI, EVI, GEMI, NDVI and NDWI indices from the WFI sensor (CBERS-4A, CBERS-4, AMAZONIA-1), aggregated to 500 m grid cells (38,957 cells) and manually labeled each year as totally burned, partially burned, or non-burned. It is openly available on ScienceDB (DOI 10.57760/sciencedb.14040).
How accurate is the burned-area classification?
Best configuration (train on three classes, then relabel TB as burned area): overall IoU 0.98/0.98/0.95 for 2020/2021/2022, but the burned-area class IoU is only 0.66/0.71/0.49. Partially burned cells are weakest, misclassified in roughly two-thirds of cases (commission error 0.64–0.73). Because train/test splits are random over adjacent cells, these figures may be somewhat optimistic.
How does it compare with the MODIS MCD64A1 burned-area product?
Errors are of similar magnitude to MCD64A1's own global validation (commission ~40%, omission ~60% for tropical savannas per Boschetti et al. 2019). For the best WFI approach, Table 11 gives BA commission errors of 22.7%/28.2%/59.3% and omission errors of 46.1%/43.2%/46% for 2020/2021/2022 — note the paper's abstract states these two error types the other way around.
Is code and data available for this paper?
Yes — the labeled dataset and the Random Forest classification script are openly available on ScienceDB (DOI 10.57760/sciencedb.14040). The full preprocessing pipeline (DOS correction, co-registration, grid extraction) is described step by step in the paper, but the availability of that pipeline's code and of the MCD64A1 comparison scripts is not explicitly stated.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–14 appear only as captions in the extract; claims resting on the distribution plots (e.g., Figure 8 density shapes) could not be verified visually.
- The abstract CE/OE transposition and the 39,957/38,957 discrepancy were assessed from verbatim-extracted text and tables; a typeset production version should be double-checked before citing these as definitive errors.
- Supplementary materials, if any, were not included in the extract; per-simulation hyperparameter selections are therefore unverifiable.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.