Paper · 2026-09-20 · cs.LG, cs.AI, cs.CV

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

Builds a 2020-2022 tabular dataset of WFI satellite time series (blue, green, red, NIR and five spectral indices) on a 500 m grid over Chapada dos Veadeiros National Park, Brazil, with cells hand-labelled as totally burned, partially burned or non-burned. A Random Forest reaches 0.94-0.95 F1 on the best binary setup; burned-area IoU is only 0.49-0.71 for the best validation setup. Author-reported.

Random forest Time series classificationremote-sensingburned-area-mappingdatasetrandom-forest

arXiv PDF Markdown14 min read

AI-compiled by claude-haiku-5-5 on 2026-11-01, not yet reviewed by a human. Every claim cites its source; check it when a detail matters.

In plain words

Forest and savanna fires leave dark scars on satellite pictures, and someone has to find them. This paper builds a dataset for doing that automatically in one national park in central Brazil. It takes three years of pictures from Brazilian satellites, cuts the park into square cells about half a kilometre on a side, and records for each cell how the colours and vegetation indices changed over time. People then went through the pictures by hand and marked each cell as burned in full, burned in part, or not burned.

The authors trained a standard Random Forest, a collection of decision trees that vote, on these cells. The best result came from treating only the fully burned cells as "burned" and everything else as "not burned". The overall score looks high, but most cells are not burned, so the score mostly measures how well the model finds unburned ground. The score for the burned class alone is much lower. The paper is a dataset with a preliminary baseline, not a finished burned-area product.

TL;DR

Dataset and baseline, not a new method. A 500 m grid over Chapada dos Veadeiros National Park (2020 to 2022), built from WFI imagery on CBERS-4A, CBERS-4 and AMAZONIA-1, with manually labelled totally burned (TB), partially burned (PB) and non-burned (NB) cells. A Random Forest trained on the four BGR-NIR bands gets F1 of 0.92 to 0.95 for binary labels tab. 7. Burned-area IoU reaches 0.49 to 0.71 in the best validation setup tab. 8. The published labels and code are the only route to reuse; the code is not on GitHub and could not be read for this card.

Problem

Burned-area products such as MCD64A1 come from MODIS at coarse resolution, and the authors say they are less accurate in the Cerrado §1. WFI sensors on three Brazilian satellites together give frequent observations at 55 to 64 m §1 §2.2.1. No systematic burned-area mapping with these satellites existed when the paper was written §1. The paper asks whether a labelled, multi-temporal tabular dataset from these sensors can support a classifier, and how the three burn classes should be grouped §1 §4.

Method

The workflow has a raster stage, a vector stage, labelling, and a classification stage fig. 2.

Study area and period. Chapada dos Veadeiros National Park (CVNP), 240,611 ha, in Goiás §2.1. Images from 1 January 2020 to 31 December 2022 §2.2.1. The analysis bounding box is the park plus a 10 km buffer §2.2.4.
Imagery. Level-4 WFI images (BGR and NIR bands) were downloaded from the INPE catalogue with the Cbers4asat package. The query used 100% cloud cover, which the authors say is not the cloud cover of the study area §2.2.1. Of 382 downloads, 147 were dropped for clouds, insufficient swath coverage or duplicate dates, leaving 235 §2.2.1 §3.1. In 2020 only CBERS-4A images were used; in 2021 CBERS-4A and AMAZONIA-1; in 2022 all three satellites, with CBERS-4 images available from 9 May 2022 §2.2.1 §3.1.
Band stacking and reprojection. Single bands were stacked into one raster per image §2.2.2. Stacks in WGS-84 UTM zones 22S and 23S were reprojected to WGS-84 (EPSG:4326) with nearest-neighbour resampling §2.2.3. Images were clipped to the common bounding box §2.2.4.
Radiometric correction (dark object subtraction). Digital numbers are converted to radiance and the dark-object radiance is subtracted §2.2.5 eq. 1 to eq. 7. The Earth–Sun distance is taken from the Ephem library for each acquisition date, and the acquisition time is set to 13:55:00 UTC §2.2.5. Band calibration constants are in tab. 1.

$G = \frac{L_{Max} - L_{Min}}{DN_{Max} - DN_{Min}}$ eq. 1

$O = L_{Min} - G \cdot DN_{Min}$ eq. 2

$L_b = G \cdot DN + O$ eq. 3

$E_{SUN} = \pi \, d^2 \, \frac{L_{Max}}{\rho_{Max}}$ eq. 4

$L_{1\%} = \frac{0.01 \cdot E_{SUN} \cos\theta}{\pi \, d^2}$ eq. 5

$L_p = L_b - L_{1\%}$ eq. 6

$\rho = \frac{\pi \, (L_b - L_p) \, d^2}{E_{SUN} \cos\theta}$ eq. 7

Symbols as printed in eq. 1 to eq. 7: DN is the pixel digital number, DN_Max and DN_Min the largest and smallest pixel values of each input band §2.2.5 eq. 1; L_Max, L_Min, rho_Max from tab. 1; d the Earth–Sun distance; theta the solar zenith angle eq. 5. See Pitfalls on eq. 7.
Resampling and co-registration. CBERS-4A images (55 m) were resampled to 64 m with nearest neighbour, to match the other two sensors §2.2.6. Each scene was co-registered to an AMAZONIA-1 reference image with Arosics, using a grid resolution of 30 pixels and a window of 400 x 400 pixels §2.2.7. Arosics plotted 2,772 control points; not all were used §2.2.7.
Spectral indices. Five indices are computed per pixel: NDVI, GEMI, NDWI, EVI and BAI tab. 2. Their formulas are in tab. 2. For EVI the paper gives C1 = 6, C2 = 7.5 and L = 1 tab. 2. The authors chose these because BAI targets burned areas, NDVI and EVI are general vegetation indices, GEMI is a non-linear vegetation index, and NDWI was used for water §2.2.8.
Gridding. A regular grid of 500 m cells, the MODIS resolution taken as the minimum mapping unit, has 239 columns and 163 rows §2.2.9. For each scene, each cell takes the zonal mean of each of the nine layers (four bands, five indices) over the pixels inside it §2.2.10.
Temporal replication. Each cell's observations are placed on a daily calendar and every day without an image takes the value of the last earlier observation (forward fill) §2.2.11 fig. 3. The authors say this is chosen because burned areas do not necessarily occur between two observations §2.2.11. The leap day 29 February 2020 is removed §2.2.12. The nine layers are concatenated in the order of tab. 3.
Labelling. Burned cells were digitised by hand in QGIS 3.22.12 using an R(R)NIR(G)B(B) false-colour composite, in which burned areas appear dark §2.2.13 fig. 4. Each cell is labelled TB, PB or NB, and the burn date of each year is stored in date columns §2.2.13 §2.2.14 tab. 4. The labelling is per cell, not per pixel §2.2.13.
Annual datasets. For each year, the four BGR-NIR bands alone form the annual classification set: 4 x 365 = 1,460 daily columns plus label columns §2.2.15 tab. 4. The annual indices are not used in the classification experiments §2.2.15 §4.
Classification. Random Forest from scikit-learn §2.2. For each year, hyperparameters are chosen by GridSearchCV over the number of trees and the maximum depth §2.2.16. Three relabelling approaches are tested, each with a balanced training sample tab. 5:
- 1st: TB as BA; PB and NB as NB.
- 2nd: TB and PB as BA; NB as NB.
- 3rd: TB, PB and NB as three classes.
Each approach is repeated 200 times with a new random draw of samples per run §2.2.16 tab. 5.

Validation. The test set is 30% of the samples §2.2.17. Precision, recall, F1 and IoU are computed per run; the paper reports mean and standard deviation §2.2.17. Seven validation approaches map the three classifiers onto binary or three-class scores, with post-classification relabelling tab. 6. The 5th validation approach is 3rd-classifier output with TB reclassified as BA and PB and NB as NB tab. 6 §2.2.17.

Pseudocode

# One Monte Carlo run of the classifier and validation  [§2.2.16-17]
# Input:  X (cells, features), y in {TB, PB, NB}, relabel for one approach
# Output: precision, recall, F1, IoU for this run

def one_run(X, y, relabel, frac_train=0.7):
    # Map the three labels to the classes of this approach  [tab. 5]
    labels = relabel(y)
    # Balanced sample: per class, n = frac_train * size of smallest class
    # [tab. 5] [§2.2.16]; how the 30% test set is drawn is not stated
    train_idx, test_idx = balanced_split(labels, frac_train)
    # Tune trees and max depth by grid search  [§2.2.16]
    model = fit_grid_search(
        X[train_idx], labels[train_idx],
        n_estimators=[100, 200, 300],   # [§2.2.16]
        max_depth=[3, 5, 7],            # [§2.2.16]
    )
    # Score on the held-out 30%  [§2.2.17]
    return scores(model, X[test_idx], labels[test_idx])

External comparison. Monthly MCD64A1 products are merged into one yearly raster of all burned pixels §2.2.18. Two reference maps per year are built from the labels (TB as BA; TB and PB as BA) §2.2.18. One of the 200 runs of the 5th approach is picked at random per year and compared to MCD64A1 with IoU, commission error and omission error §2.2.18 tab. 11.

Hyperparameters

Name	Value	Where	Notes
study area	Chapada dos Veadeiros NP, 240,611 ha	§2.1
buffer around park	10 km	§2.2.4	Brazilian rule cited
years	2020, 2021, 2022	§2.2.1	each year modelled separately §2.2.14
images used	235 of 382 downloads	§2.2.1 §3.1	cloud threshold for the filter: not stated
query cloud cover	100%	§2.2.1	authors say this is not the study-area cloud cover
raw spectral source	WFI on CBERS-4, CBERS-4A, AMAZONIA-1, L4	§2.2.1
reprojection	nearest neighbour, EPSG 32722/32723 to 4326	§2.2.3
resampling	CBERS-4A 55 m to 64 m, nearest neighbour	§2.2.6
co-registration grid resolution	30 pixels	§2.2.7	Arosics
co-registration window	400 x 400 pixels	§2.2.7
EVI constants	C1 = 6, C2 = 7.5, L = 1	tab. 2
grid cell size (MMU)	500 m x 500 m	§2.2.9
grid size	239 columns x 163 rows	§2.2.9	38,957 cells §3.3
daily columns per layer	1,096 (text); 1,095 implied by tab. 3	§2.2.12 tab. 3	leap day removed
annual classification columns	1,460 (4 bands x 365 days)	§2.2.15
n_estimators	100, 200, 300	§2.2.16	GridSearchCV grid
max_depth	3, 5, 7	§2.2.16	GridSearchCV grid
Monte Carlo runs	200 per approach	§2.2.16 tab. 5
train / test split	70% / 30%	§2.2.16 §2.2.17
training size per class	see tab. 5	tab. 5	equals 70% of smallest class count
validation approaches	7	tab. 6
MCD64A1 runs compared	1 random run per year	§2.2.18
Random Forest criterion, max_features, class_weight, min_samples, CV folds, seed	not stated

Training size per class, read from tab. 5:

Approach	2020	2021	2022
1st	1,505	1,726	1,457
2nd	3,254	3,432	3,659
3rd	1,505	1,705	1,457

These match 70% of the smallest class count for each year and approach, using the cell counts in §3.3. The authors do not describe the draw.

Results

Author-reported. Setup: CVNP, 2020 to 2022, one Random Forest per year and approach, 200 Monte Carlo runs, 30% test §2.2.16 §2.2.17.

Cell counts in the labelled grid (38,957 cells, §3.3):

Year	TB	PB	NB	Source
2020	2,151 (5.52%)	2,498 (6.41%)	34,308 (88.06%)	§3.3
2021	2,467 (6.33%)	2,437 (6.25%)	34,053 (87.41%)	§3.3
2022	2,082 (5.34%)	3,146 (8.07%)	33,729 (86.58%)	§3.3

Test-set scores (mean over 200 runs, tab. 7):

Approach	Metric	2020	2021	2022	Source
1st	F1	0.94	0.95	0.92	tab. 7
1st	Precision BA	0.94	0.94	0.90	tab. 7
1st	Recall BA	0.94	0.96	0.95	tab. 7
2nd	F1	0.91	0.93	0.89	tab. 7
2nd	Precision BA	0.94	0.94	0.88	tab. 7
2nd	Recall BA	0.87	0.91	0.90	tab. 7
3rd	F1	0.79	0.79	0.74	tab. 7
3rd	Precision TB	0.82	0.81	0.75	tab. 7
3rd	Recall TB	0.79	0.78	0.76	tab. 7
3rd	Precision PB	0.71	0.70	0.62	tab. 7
3rd	Recall PB	0.69	0.70	0.66	tab. 7
3rd	Precision NB	0.84	0.87	0.87	tab. 7
3rd	Recall NB	0.89	0.90	0.80	tab. 7

Validation IoU over the full annual datasets (mean over 200 generalisations, tab. 8). Burned-area (BA) IoU is the number to read, because NB is most of the grid:

Validation approach	IoU 2020	IoU 2021	IoU 2022	BA IoU 2020	BA IoU 2021	BA IoU 2022	Source
1st (clf 1, TB as BA)	0.94	0.94	0.90	0.48	0.52	0.34	tab. 8
2nd (clf 1, TB+PB as BA)	0.94	0.95	0.91	0.58	0.68	0.52	tab. 8
3rd (clf 2, TB as BA)	0.91	0.90	0.83	0.52	0.39	0.24	tab. 8
4th (clf 2, TB+PB as BA)	0.94	0.95	0.89	0.66	0.69	0.54	tab. 8
5th (clf 3, TB as BA)	0.98	0.98	0.95	0.66	0.71	0.49	tab. 8
6th (clf 3, TB+PB as BA)	0.90	0.91	0.82	0.53	0.56	0.41	tab. 8
7th (clf 3, three classes)	0.85	0.85	0.76	TB 0.75 / 0.75 / 0.66	PB 0.76 / 0.73 / 0.73	NB 0.99 / 0.99 / 0.96	tab. 8

Comparison with MCD64A1 (one run per year, tab. 11):

Reference / product	IoU	BA IoU	NB IoU	CE (BA)	OE (BA)	Source
5th approach 2020 vs MCD64A1 2020	0.867	0.464	0.932	0.227	0.461	tab. 11
5th approach 2021 vs MCD64A1 2021	0.811	0.463	0.895	0.282	0.432	tab. 11
5th approach 2022 vs MCD64A1 2022	0.798	0.301	0.892	0.593	0.460	tab. 11
TB as BA 2020 vs MCD64A1 2020	0.876	0.453	0.935	0.493	0.189	tab. 11
TB+PB as BA 2020 vs MCD64A1 2020	0.873	0.551	0.927	0.166	0.379	tab. 11

The paper's headline error figures are in the abstract. See Pitfalls.

Confusion matrices for the 7th approach, rows true and columns predicted tab. 10:

Year	True	Pred TB	Pred PB	Pred NB
2020	TB	1,819	296	36
2020	PB	317	1,948	233
2020	NB	404	3,230	30,674
2021	TB	2,005	434	28
2021	PB	297	1,933	207
2021	NB	73	3,162	30,818
2022	TB	1,767	295	20
2022	PB	553	2,353	240
2022	NB	1,108	5,916	26,705

Commission and omission error, 7th approach tab. 9: TB CE 0.284 / 0.156 / 0.484, OE 0.154 / 0.187 / 0.151; PB CE 0.644 / 0.65 / 0.725, OE 0.22 / 0.206 / 0.252; NB CE 0.0087 / 0.0076 / 0.0096, OE 0.106 / 0.095 / 0.208 (2020 / 2021 / 2022).

Paper vs code

Code not retrieved. The paper says the Random Forest script is in a ScienceDB repository §2.2.15 §3.2, and the data are at the ScienceDB DOI 10.57760/sciencedb.14040 [data availability statement, p. 25]. The DOI redirects to a scidb.cn dataset page. That page returned only its title, with no file list, authors or licence, so the script could not be read, and no code licence is known. No code is recorded on this card.

Checked and not used: a GitHub repository search for "burned area WFI cerrado" returned two MIT-licensed repositories, agojunior/BurnedAreaDetection-Cerrado-WFI and AOlivr/wfi-time-series-burned-areas-cerrado. Neither owner is an author, and the paper does not link either, so they are not the official code.

Paper discrepancies that a reader can check from the text alone, without the script:

eq. 7 with eq. 5 and eq. 6 reduces to a constant; see Pitfalls.
The GEMI formula in tab. 2 prints "(1 - 0.25)" where the usual form has gamma in that term; see Pitfalls.

Pitfalls

Surface reflectance equation. Substituting eq. 6 into eq. 7 makes the numerator pi (L_1%) d^2, so rho = 0.01 for every pixel. That cannot be what produced the data. A standard dark-object-subtraction reflectance uses the subtracted radiance L_p in the numerator. The code is not readable, so which form was used is not stated. Do not implement eq. 7 as printed without checking.
Grid cell count. §2.2.9 says 39,957 cells. 239 x 163 = 38,957, and the per-year class counts in §3.3 sum to 38,957. Use 38,957.
Abstract error numbers. The abstract gives BA omission 22%, 28%, 59% and commission 46%, 43%, 46%. tab. 11 gives BA commission 0.227, 0.282, 0.593 and omission 0.461, 0.432, 0.460 for the 5th approach. The abstract appears to swap the two labels. Quote tab. 11 and state its definitions.
IoU in the text does not match Table 8. §3.3 gives 5th approach IoU 0.97, 0.97, 0.95 and 1st approach 2022 as 0.89, and 2nd approach 2022 as 0.90. tab. 8 gives 0.98, 0.98, 0.95; 0.90; and 0.91. Use the table.
IoU may be in-sample. §3.3 says Table 8 is "generalizing the models to the full annual datasets". The full dataset includes the 70% training cells. Only tab. 7 is clearly on held-out cells. The paper does not say whether training cells were removed for IoU.
Overall IoU is dominated by NB. NB is 86% to 88% of cells §3.3, so overall IoU is high even when burned-area IoU is 0.3 to 0.7. The "5th approach is best" claim in §4 rests on overall IoU and on BA IoU that is only comparable to the 4th approach in 2020 and 2021 tab. 8.
Approach choice is made on the reported numbers. The 5th approach was selected after comparing all seven on the same annual datasets §3.3. A cleaner test would pick it on one split and report another.
Std is described but not reported. §2.2.17 says mean and standard deviation were computed; the tables give means only. Run-to-run spread is unknown.
Test-set draw is unclear. tab. 5 gives training sizes equal to 70% of the smallest class, and §2.2.17 says 30% is for testing. The paper does not say whether test cells are drawn from the same balanced pool, from the remainder, or stratified. Cells are also adjacent 500 m squares with correlated burn status, and a random cell split can leak neighbouring information. The paper does not discuss spatial autocorrelation.
Missing Random Forest settings. Split criterion, max_features, min_samples_leaf, class_weight, CV folds and random seed are not stated. The paper says it did not check which grid values were chosen §3.2.
Replication before first image. Forward fill needs a starting value. §2.2.11 does not say what happens before a cell's first observation in a year.
Column counts conflict. §2.2.12 says 1,096 columns per layer, then removes 29 February 2020, which leaves 1,095. tab. 3 gives ranges of 1,095 per layer. tab. 4 puts the date_20 column at index 9856, which is the last index in tab. 3's NDWI range. Check indices against the dataset file, not the text.
Band order. §2.2.2 says the stack is "RGB NIR", while fig. 2 labels B1 as Red, B2 as Green, B3 as Blue, and tab. 3 lists blue first. Confirm band order from the file metadata.
GEMI formula. tab. 2 prints the second term as gamma x (1 - 0.25). The usual GEMI uses gamma x (1 - 0.25 gamma). The paper does not say which was computed.
Label name typo. §2.2.13 writes "partially burned (TB)". It means PB.
Cloud threshold. The image filter is described only as "clouds" with the query at 100% §2.2.1. The numeric cut-off, and how swath coverage was judged, are not stated.
Sensor mix changes by year. 2020 uses only CBERS-4A. 2021 adds AMAZONIA-1. CBERS-4 arrives on 9 May 2022 §2.2.1 §3.1. Year effects therefore mix sensor effects, which the paper names but does not separate §4.
Labels are one-off manual work. The paper does not report how many people labelled the cells, whether they checked each other, or how they handled ambiguous cells. Treat labels as the authors' interpretation of fig. 4.
Only BGR-NIR is used in the experiments. The five indices are in the dataset but not used in §2.2.15 or §3. Results for the indices are not reported.
Three years, one park. Each year is a separate model and the study covers one 240,611 ha park. Nothing here shows transfer to other biomes.
Use of MCD64A1 as reference. MCD64A1 is 500 m, while the labels are built from 64 m WFI pixels §3.4. The paper treats MCD64A1 as a comparison product, not ground truth, and reports that the two disagree.
Licence of the dataset. The paper is CC BY 4.0. The ScienceDB dataset licence is not stated in anything readable here. Do not treat the dataset as licensed for reuse without checking that page.

Lineage

Builds on: standard dark-object subtraction (Chavez-Jr, 1988) §2.2.5; Arosics co-registration (Scheffler, 2017) §2.2.7; scikit-learn Random Forest (Pedregosa et al., 2011) §2.2; and the WFI sensor harmonisation work of Oldoni (2022) §1 §3.1.
Compared against: MCD64A1 (Giglio et al., 2018) §2.2.18 tab. 11.
Superseded by: not stated.

Source and license

Adapted from “A multi-temporal dataset for mapping burned areas in the Brazilian Cerrado using time series of remote sensing imagery” by Alisson Cleiton de Oliveira, Thales Sehn Körting (arXiv:2609.23687), used under CC BY 4.0. Changes: condensed and restructured for implementation, with tables, equations and algorithms taken from the paper and its code, each cited to its source. Not affiliated with or endorsed by the authors.