Fig 1.
Example of a Sentinel-2 Normalized Burn Ratio (NBR) time series.
The example corresponds to a single pixel (20 × 20 m) in Kalimantan, where the reference month being processed is September. The pre- and post-fire compositing windows used in the monthly burned-area processing chain are indicated.
Fig 2.
Monthly versus annual Sentinel-2 pre- and post-fire composites over an area in Central Kalimantan.
Comparing Sentinel-2 pre- and post-fire monthly composites (August-November 2019) with the annual composites published previously [30]. Sentinel-2 composites are displayed in false colours (RGB: short-wave infrared, band 11; near infrared, band 8; blue: red, band 4). The images reveal large burn scars, visible as areas that have transitioned from green to dark brown/red tones. The Monthly sequence reveals the progression of the burn scars.
Fig 3.
Location of training and validation points labelled as ‘Burned’ and ‘Unburned’.
The map above shows the 2,343 training points used to train the Random Forest. The map below shows the 1,042 reference points used to validate the burned area map.
Fig 4.
Persistence of burn scar visibility leading to repeated monthly detection by Sentinel-2 imagery.
Pre- and post-fire Sentinel-2 monthly composites, and corresponding burned-area classifications for February, March, April, and May 2019 are shown over an area (Pulau Rupat) in Riau province, Sumatra where peatland burn scars remained visible several months after fire. Pixels that were detected as ‘Burned’ for the first time are shown in red, while pixels that were repeatedly detected as ‘Burned’ are shown in orange.
Fig 5.
Illustrative representation of validation points used for the monthly burned-area layer.
From the original 1,168 points, we adjusted the sample to ensure proper stratified random sampling in the monthly burned-area layer in 2019.
Fig 6.
Area burned by number of times between 2019 and 2024.
Fig 7.
Spatial distribution of the burned area fraction (2019–2024) across Indonesia, mapped at a 1 km resolution.
Each panel represents the burned area fraction for a specific year, calculated as the proportion of the pixel area affected by fire during that year. The color gradient from yellow to red indicates the burned area fraction, ranging from low (0%) to high (100%). The maps reveal significant spatial and interannual variability in burned area patterns. For this figure, the 20-m burned area layer was aggregated to a 1-km burned area fraction to allow the visualization at the country level. A 20-m map has too much detail to be clearly shown over large areas.
Fig 8.
Total burned area by region (left) and province (right) in Indonesia from 2019 to 2024.
Burned area is shown in million hectares (Mha), with bars stacked by year to illustrate interannual variability. The highest levels of fire activity occurred in Kalimantan, Bali & Nusa Tenggara, and Sumatra, with the provinces of Nusa Tenggara Timur, Kalimantan Tengah, Papua Selatan, and Sumatera Selatan accounting for 52.2% of the national total across the six-year period. Peak fire years (2019 and 2023) dominate most provinces, while lower activity is seen in 2021 and 2022.
Fig 9.
Indonesia-wide cumulative burned area (in million hectares, Mha) by month for years 2019 to 2024.
This graph compares results from this study (left panel, red lines) and the MCD64A1 burned area product (right panel, black lines). Each line represents the cumulative burned area for a given year, with final annual totals indicated on the right. The sharp increase in burned area from June to October reflects the seasonal pattern typical of fire activity in Indonesia, driven by dry conditions during these months. The consistent underestimation by MCD64A1 highlights the improved detection capacity of this study’s approach.
Fig 10.
Monthly burned area (top panel) and climate indices (bottom panel) for Indonesia from January 2019 to December 2024.
The top panel shows total burned area across all provinces (in million hectares, Mha). The bottom panel displays the Indian Ocean Dipole (IOD; black line, left axis, °C) and Oceanic Niño Index (ONI; blue line, right axis, °C), representing major climate anomalies. Peaks in burned area during late 2019 and mid to late 2023 align with positive IOD and El Niño conditions, while reduced burned area between 2020 and 2022 corresponds to neutral or negative phases of both indices. Vertical lines mark the start of each calendar year.
Table 1.
Accuracy assessment of each of the three burned-area maps performed in seven Indonesian provinces (87.60 Mha) targeted for peatland restoration. The accuracy metrics were estimated with a 1,042 points randomly distributed using stratified sampling. The reported metrics are (1) the overall accuracy (OA), the user accuracy (UA), and the producer accuracy (PA) with their 95 % confidence intervals.
Fig 11.
Confusion matrix of burn-month detection from processing chain versus visual interpretation.
This matrix compares the month of detection assigned by the monthly burned-area processing chain (NRT, y-axis) with the burn month determined by visual interpretation of Sentinel-2 imagery (x-axis) for 259 reference sites in 2019. The matrix shows the number of sites detected as “burned” in each month by the processing chain versus the month confirmed through visual inspection. High counts along the diagonal indicate strong agreement between the processing chain and visual interpretation. Off-diagonal values reflect either persistent visibility of burn scars or timing discrepancies of detection.
Fig 12.
Cumulative burned area as a function of burned patch size for each year from 2019 to 2024.
The red line shows results from this study, while the black line represents the MODIS MCD64A1 burned area product. For each year, cumulative burned area (in million hectares, Mha) is plotted against individual burned patch sizes (in hectares, ha, logarithmic scale). The x-axis represents burned patch size on a logarithmic scale, allowing for a clear visualization of both small and large patches, which vary over several orders of magnitude. Across all years, this study consistently maps a larger cumulative burned area and detects more burned patches across a wide range of sizes compared to MCD64A1. The differences are especially pronounced for large patches in major fire years (2019 and 2023), but are also evident for small, burned patches, particularly in low-fire years (2020–2022), indicating that MCD64A1 underestimates both small and large fires in Indonesia.
Table 2.
Tests statistics with respect to differences in burned-area scar size frequency distributions for Sentinel, MODIS, and official maps.