Rapid extraction of forest burned areas using Sentinel-2 satellite imagery on the PIE-engine platform

2.1. Study area and data sources

Liuhe Village lies in the eastern mountainous region of Xugu Town, Xinzhou District, Wuhan (115°02'01"E-115°02'37"E, 30°53'55"N-30°54'13"N). The study area covers an area of 1,200 hectares, including 70.87 hectares of farmland and 516.2 hectares of forested land. This region is characterized by a subtropical monsoon climate with abundant rainfall and high temperatures during the synchronous rainy and hot seasons. The terrain is dominated by low hills and mountains, and the vegetation is mainly composed of highly flammable pine trees and understory grasses. On October 4, 2019, extreme drought and elevated temperatures triggered a forest fire near Liuhe Village, and the fire danger index continued to reach its maximum level (red alert). Field surveys conducted on December 21, 2019 used a GPS data collector (Unistrong MG758) to record the boundaries of burn scars. The integration with satellite imagery during post-processing confirmed that the burned area was 106.25 hectares (Fig. 1).

This study used Sentinel-2 satellite data accessed by the PIE-Engine platform (https://engine.piesat.cn/engine/home). The Sentinel-2 mission, operated by the European Space Agency, used twin satellites (Sentinel-2A and 2B) to achieve a combined 5-day revisit cycle. Each satellite was equipped with a Multispectral Imager (MSI) that can capture13 spectral bands in the visible, near-infrared (NIR), and shortwave infrared (SWIR) regions, with a spatial resolution of up to 10 m. We used the Level-2A Bottom-of-Atmosphere reflectance products (Dataset ID: “S2/L2A”) provided by PIE-Engine.

Fig. 1Location of the study area. Source: National Geomatics Center of China, Tianditu imagery base map [GS(2025)1508], retrieved March 2025, from https://www.tianditu.gov.cn

2.2. Advantages of data preprocessing and platform computation

One of the core advantages of implementing this method on the PIE-Engine platform is the use of its cloud-native parallel computing architecture. Unlike traditional desktop remote sensing software (such as ENVI and ArcGIS) that requires downloading, local storage, and sequential processing of large satellite image datasets (a process often taking minutes to hours), PIE-Engine performs these computations in parallel on its servers. This architecture allows for near-instantaneous access, preprocessing, and analysis of Sentinel-2 imagery directly in the web browser, which is crucial for achieving the fast-processing times. A web-based application was designed to visualize results, featuring automated threshold calculation, manual threshold adjustment, and binary map generation. The workflow is shown in Fig. 2.

Sentinel-2 Level-2A images are provided by PIE-Engine with atmospheric correction. For each image including the study region and period, the cloudMask function is used for cloud removal. False-color synthesis was generated by assigning SWIR (Band 12), NIR (Band 8), and red (Band 4) bands to the RGB channels respectively to improve the visibility of burned scars. All available cloud-free images were fused and clipped to the study area boundary, and resampled to are solution of 10 m. Fig. 3 shows the pre-fire and post-fire false-color synthesis.

Fig. 2Flowchart of the methodology

Fig. 3Sentinel-2 false color composite images a) before and b) after the fire

a)

b)

2.3. Spectral index calculation

For the pre- and post-fire image synthesis, three spectral indices are calculated as follows:

1) Normalized Difference Vegetation Index (NDVI) [37]:

2) Normalized Burn Ratio (NBR) [38]:

3) Normalized Burn Ratio 2 (NBR2) [39]:

where ${\rho }_{B4}$, ${\rho }_{B8}$, ${\rho }_{B11}$, and ${\rho }_{B12}$ represent the BOA reflectance values of Sentinel-2 Band 4 (Red, 665 nm), Band 8 (NIR, 842 nm), Band 11 (SWIR1, 1610 nm), and Band 12 (SWIR2, 2190 nm), respectively.

The difference indices (dNDVI, dNBR, dNBR2) were calculated by subtracting the pre-fire index value from the post-fire index value for each pixel [40], with an output resolution of 10 m.

2.4. Threshold extraction and area calculation

The Otsu algorithm is used to obtain an adaptive threshold for segmenting burned areas from each difference index image [41]. The algorithm finds the threshold $T$ that maximizes the inter-class variance between the foreground (burned) and background (unburned) pixel classes:

where $w_0$ and $w_1$ are the probabilities of the two classes separated by threshold $T$; $u_0$ and $u_1$ are the mean intensities of the two classes, and $u$ is the global mean intensity of the difference image.

Although the Otsu threshold ($T_{Otsu}$) provides a statistically robust starting point, errors in residual commission (unburned areas classified as burned) or omission (missed burned areas) may occur in heterogeneous landscapes. To address this, this tool includes a manual adjustment interface. Users can fine-tune the final threshold ($T_{final}$) by applying a constant offset ($\mathrm{\Delta }$) to $T_{Otsu}$: $T_{final}=T_{Otsu}+\mathrm{\Delta }$, where $\mathrm{\Delta }$ can be increased or decreased in steps of 0.05 based on visual inspection of the preliminary binary map and local knowledge. This simple and effective post-Otsu adjustment allows users to account for spectral changes at specific locations. The core adaptive thresholding framework (Otsu + manual calibration) is universal and can be reused to detect other anomalies (such as floods, and deforestation) by replacing the input spectral indices with relevant ones (such as NDWI for water).

The burned area ($A$) in hectares is calculated as: ${A=N_{burnedpixels}\mathrm{*}\frac{10\mathrm{*}10}{10000}=N}_{burnedpixels}\mathrm{*}0.01$, where $N_{burnedpixels}$ is account of pixels classified as burning.

2.5. Web-based user interface

The entire workflow runs locally in the PIE-Engine platform. A browser-accessible interface is built with the UI components of the platform: the InputNumber widgets for coordinate and threshold offset ($\mathrm{\Delta }$), the DateSelect components for defining pre-fire and post-fire time ranges, the Select dropdowns for selecting the spectral index, and a Button for trigger processing. All widgets are nested in a Panel container for a responsive layout. After execution, the application will display an interactive map with a binary burned area mask and output the calculated area in hectares.

2.6. Accuracy assessment

The main accuracy of the burned area range is evaluated by comparing the remotely sensed estimates ($S_i$) with the ground-surveyed reference area ($S_{ref}$ = 106.25 ha). The area accuracy $P_{area}$ (%) was calculated as follows:

To enhance the quantitative evaluation, we have added pixel-level classification metrics from a stratified random sample of points ($n=$500). This sample is used to calculate the confusion matrix and obtain the following standard classification indicators:

Accuracy (User Accuracy): Measures the reliability of the classification, that is, the proportion of predicted burned pixels that are actually burned:

Recall (Producer Accuracy): Measures the completeness of the classification, that is, the proportion of actual burned pixels that are correctly identified:

F1-Score: the harmonic mean of accuracy and recall, providing a single balanced metric:

Cohen’s Kappa Coefficient ($\kappa$): Measures the consistency between classification and reference data, exceeding accidental expectations:

where $TP$ (True Positives) is the number of pixels correctly classified as burned; $FP$ (False Positives) refers to the number of pixels incorrectly classified as burned (commission error), and $FN$ (False Negatives) is the number of burned pixels incorrectly classified as unburned (omission error). $P_0$ is the observed agreement (overall accuracy), and $P_e$ is the expected accidental consistency calculated based on the marginal sum of the confusion matrix.

In addition, a comparative experiment was conducted to place the performance of this index-based method in the background. A Random Forest (RF) classifier is implemented on the same PIE-Engine platform using the pre-fire and post-fire reflectance values of bands B2, B3, B4, B8, B11, and B12 as features. The model is trained on a separate set of labeled data from the same event. Its performance (Precision, Recall, F1, Kappa) and computational execution time are compared with the dNBR+Otsu method to highlight the trade-offs between accuracy, complexity, and operational efficiency.

3.1. User interface and processing efficiency

The web application developed on the PIE-Engine platform demonstrated efficient burned area extraction through its intuitive interface and robust processing algorithms. As shown in Fig. 4, there is a parameter configuration panel (Fig. 4(a)) on the left sidebar of the interface. It is used to define the area coordinates, temporal ranges, and processing thresholds. The system automatically selects optimal pre-fire and post-fire Sentinel-2 scenes in the user-specified date range. The calculation results include the adaptive thresholds and area metrics, which are dynamically shown in the lower left panel (Fig. 4(b)). The central mapping area (Fig. 4(c)) visualizes the results, and the layer management module (Fig. 4(d)) allows for flexible switching of layers. This responsive design ensures full functionality in devices without the need to specialized software.

3.2. Automatic burned area identification

Fig. 5 shows the adaptive-threshold burned area maps generated within about 2 s. There are significant performance differences between the indicators. The dNBR index produces the most spatially coherent results with minimal noise (Fig. 5(b)). The dNDVI output contains a significant amount of noise and omission errors, particularly misclassifying terrain shadows in northwestern regions due to its known sensitivity to illumination variations, atmospheric interference, and solar geometry effects (Fig. 5(a)). The dNBR2 index performs moderately and effectively distinguishes agricultural changes in eastern farmlands (Fig. 5(c)). Common challenges include false alarms caused by human interference near settlements, seasonal errors in agricultural classification, and incorrect delineation of water-vegetation boundaries during hydrological fluctuations. These limitations highlight the complex interactions between spectral signatures and landscape changes in post-fire environments.

Fig. 4User interface and main functions of the application

Fig. 5Burned area maps recognized by adaptive threshold of: a) dNDVI; b) dNBR; c) dNBR2

a)

b)

c)

3.3. Quantitative accuracy and comparative evaluation

Quantitative assessment confirms the superiority of dNBR. The adaptive thresholding approach using dNBR achieve an area accuracy ($P_{area}$) of 98.22 % (108.14 ha compared to the ground truth of 106.25 ha) at an optimal threshold of 0.32 (Table 1). The pixel-level classification metrics obtained from 500 validation points provide a more detailed evaluation (Table 1). dNBR achieves the highest F1-score (0.94) and Kappa coefficient (0.92), indicating an excellent balance between precision (0.95) and recall (0.93).Although manually adjusted dNDVI can achieve a high area accuracy (99.64 %), this comes at the cost of a severe damage recall (0.71), leading to significant omission errors and a lower F1-score (0.81). The visual examination of the results (Figs. 6-7) confirms that the dNBR adaptive thresholding approach best balances the accuracy and reliability, effectively minimizing omission and commission errors while maintaining the integrity of burn scar.

To address the uncertainty of boundary delineation, a 30-meter buffer zone around the reference fire perimeter is analyzed. In this boundary region, the dNBR method maintains a high classification accuracy (Producer accuracy > 90 %), indicating robust performance in distinguishing the gradient between burned and live vegetation, although some inevitable edge pixel mixing. A Random Forest (RF) classifier is implemented on the PIE-Engine for comparison. As shown in Table 1, the RF model achieves a slightly higher pixel-level accuracy (F1 = 0.96, $\kappa$ = 0.95) and area accuracy (99.66 %). However, its execution time on the cloud platform is significantly longer (about 45 s for model prediction) than the near-instantaneous processing (about 2 s) of the dNBR+Otsu method. This comparison highlights the trade-off empirically: although advanced machine learning can provides light accuracy gains, this spectral index method achieves an excellent balance between computational efficiency, operational simplicity, and “good-enough” accuracy for rapid responses.

Fig. 6Accuracy evaluation result of burned area after threshold adjustment

3.4. Ethical, social, and practical implications

The deployment of automated remote sensing tools for forest fire analysis is of great importance. First, regarding transparency and algorithmic decision-making, the proposed methodology (Otsu on dNBR) is fully transparent and deterministic. This contrasts with “black-box” deep learning models and is crucial for generating evidence that can be carefully examined and interpreted in policy-making or legal liability assessments. Then environmental and privacy considerations must be balanced. Although satellite monitoring is indispensable for environmental protection and disaster management, its use must be balanced with privacy regulations, especially when capturing private property with high-resolution data. This tool uses medium-resolution Sentinel-2 data (10 m), which can alleviate personal privacy while still being suitable for area-based damage assessment. Finally, ensuring fair access to such technology is a key practical issue. The design of this tool on the cloud-based PIE-Engine platform directly addresses this only using a web browser, thereby eliminating the need for expensive computing hardware or specialized software licenses. This is particularly beneficial for forestry departments or researchers in regions with limited technical infrastructure, reducing the barrier to advanced geospatial analysis.

Fig. 7Burned area map recognized by adjusted threshold (0.21) for dNDVI