Burned Area Mapping Research Articles

Burned-area sub-pixel mapping based on spatial–spectral information at super-pixel scale for multi-spectral image

ABSTRACT Due to the complex environments of burned areas and limitations of hardware devices, the collected multi-spectral image (MSI) is sometimes with many mixed pixels to hinder the accurate mapping of burned areas. To solve this problem, sub-pixel mapping (SPM) technology has been applied to handle with these mixed pixels to map burned areas. However, the spatial–spectral information of burned areas used by SPM is usually constructed in a specified rectangular local window, and the number of spectral bands utilized by SPM is also little, affecting the mapping accuracy of burned areas. To improve the mapping accuracy of burned areas, we propose burned-area SPM based on spatial–spectral information at super-pixel scale for multi-spectral image (SSIASC). In SSIASC, super-pixels representing the burned areas with irregular distribution are obtained by interpolation and then segmentation of the original coarse MSI. The extended random walker algorithm is then used to calculate the spatial correlation in super-pixels to obtain spatial term, and at the same time the normalized model is constructed to calculate all the spectral bands in super-pixels to yield spectral term. Next, the two terms are integrated to produce the objective term with spatial–spectral information at super-pixel scale. Finally, particle swarm optimization is employed to optimize the objective term to derive the burned-area mapping result. Experimental results on the two burned-area MSIs show that the proposed SSIASC produces the better results than the state-of-the-art SPM methods.

Mapping wildfires in Canada with Landsat MSS to extend the National Burned Area Composite (NBAC) time series back to 1972

Background Satellite imaging has improved burned area mapping; however, few studies have taken advantage of the Multi-Spectral Scanner (MSS) in early Landsat satellites, which started acquiring data 10 years earlier than Thematic Mapper (TM). Aims To expand Canada’s National Burned Area Composite (NBAC) annual time series back to 1972 using MSS data and report annual statistics and national trends for 1972–2022. Methods Pre- and post-fire image composites were created using an improved collection of MSS data available from the Google Earth Engine. A Normalized Difference Vegetation Index (NDVI) difference image was adaptively thresholded to extract burned areas, which were then vectorised. To assess accuracy, MSS fire polygons were compared with TM in a year of overlap. Key results Compared with TM, MSS polygons overestimated burned area by 5.6% when the relativised differenced NDVI was used, with significant upward trends for number of fires > 200 ha, fire season length and mean duration of fires. Conclusions MSS is a valuable data source for retrospective mapping of boreal and temperate forest fires where data from finer-resolution sensors are lacking. Implications After the addition of MSS-mapped fires, NBAC is the longest satellite-based time series of annual burned area from individually mapped fires in the world.

Unsupervised Bushfire Burn Severity Mapping Using Aerial and Satellite Imagery

Abstract. It is critical to assess bushfire impact rapidly and accurately because bushfires play a significant role in forest degradation and present a threat to ecosystems and human lives. Over the past decades, several supervised algorithms of burn severity mapping have been proposed, facing the significant drawback of time-consuming labeling. Moreover, there is no robust framework for burn severity mapping through fusing multi-sensor, multi-resolution, and multi-temporal remote sensing imagery from satellite and aerial platforms. Therefore, this paper presents an unsupervised two-step pipeline: processing 2D data followed by 3D data for burn severity mapping, both of which are acquired from either aircraft or satellites. For the 2D data processing, our proposed unsupervised burned area detection (UsBA detection) model enhances burned area mapping accuracy by integrating Ultra-High Resolution (UHR) aerial imagery with bi-temporal medium-resolution PlanetScope imagery, using a Segment Anything Model (SAM)-assisted UNetFormer (pre-trained on the target-style public dataset – LoveDA Rural) for refinement. The model demonstrates superior burned area segmentation, evidenced by improved evaluation metrics calculated from labeled test sites. For the 3D analysis, the burned areas extracted from 2D processing are further assessed using pre- and post-event airborne laser data. We implement a voxel-based workflow, including necessary steps such as ground filtering through Superpoints in RANSAC Planes (SiRP) method and biomass change analysis. The results indicate that the 3D branch provides a reliable lower bound of the actual damage map, because the vegetation growth between two measurements remains, in essence, undetected. The proposed framework offers a more accurate and robust solution for burn severity mapping utilizing combined 2D and 3D data, evaluated on a multi-source dataset from a real bushfire event that occurred in Bushland Park, South Australia.

Enhancing Burned Area Mapping Accuracy: Integrating Multi-temporal PCA with NDVI Analysis

Forested lands in the west coast of Turkey, with their similarity to Mediterranean forests, are often found to be highly susceptible to wildfires, necessitating the development of a forest management program to refine and quantify forest fires and their impacts on the environment. In light of this fact, a multi-temporal approach combining Principal Component Analysis (PCA) and Normalized Difference Vegetation Index (NDVI) analysis derived from Sentinel-2 imagery is suggested in the current study. Through PCA of carefully selected bands of Sentinel-2, both recent and historic fire impacts are attempted to be captured. It was found that the first two principal components (PC1 and PC2) predominantly describe landscape characteristics, while the third and fourth components (PC3 and PC4) have high abilities in detecting burn scars. It is worth noting that an increase in the ability to detect burn scars was observed with the inclusion of NDVI and its difference in time (dNDVI) within the PCA process. A high effectiveness level in distinguishing burnt areas from unburnt landscapes was presented by the multi-temporal PCA approach, particularly with dNDVI integration. PC2 and PC3, especially with dNDVI integration, are found to be strong indicative factors of burnt areas. In the classification result, accuracies of different years of fire events differed, and a high accuracy of 98.76% was found in the last fire event year of 2019. However, slight underestimation and overestimation were also observed in older fire scars. Mean accuracy, on average, for the PCA-dNDVI method was found to be higher than that of the MLC method. Furthermore, significant vegetation losses by fire, particularly by the 2019 fire incident, were realized through NDVI assessment. Although it worked well in recent fire scars, overestimating the extent in the case of burned areas from previous years was observed. The potential of multi-temporal PCA integration with NDVI for analysis in mapping burned areas at different scales in fire-prone ecosystems in western Turkey is underlined by the results of this work. Much more successful forest management and assessment strategies after fires have occurred in these ecosystems are helped to be created by this approach. Moreover, the approach is suggested to be one of the strong tools for monitoring fire induced damages across many time scales toward better understanding and management of long-term impacts caused by forest fires in the region.

Machine learning approach to burned area mapping for Southern California

ABSTRACT Accurate representation of the location and amount of burned areas is vital to the understanding of patterns and impacts of fires. Some extant burned area maps appear to have high commission errors, which lead to an overrepresentation of burned area. The primary research objective of this study was to assess whether region-specific training data used for machine learning routines improve accuracy of burned area products for western San Diego County. We used training data derived from fine-scale aerial orthoimagery to create and compare three training sets, each with a different Landsat scale sub-pixel burn threshold: 20%, 50%, or 80%. Meaning either 20%, 50%, or 80% of a 30 m × 30 m pixel had to burn for the entire pixel to be classified as burned. High-resolution orthoimagery was used for the creation of the training/testing data as well as determining which sub-pixel threshold leads to more accurate burned area representation. These training data were input into a gradient-boosted regression model. We compared the burned area product from the region-specific gradient boosted model (L-GBRM) to the three products: Monitoring Trends in Burn Severity, Fire and Resource Assessment Program, and the Landsat Burned Area product. We found >20% sub-pixel burn threshold of a Landsat pixel yielded the most accurate classification results. We used a 50% sub-pixel burn threshold for the reference data to compare the results to since the burned area associated with it is closely aligned with the high-resolution orthoimagery determined burn area. The L-GBRM was the most accurate product while also mapping the smallest area burned, suggesting that the extant products have relatively high commission errors. Using region-specific training data achieved a higher accuracy than nationwide training data. Looking at sub-pixel burn thresholds for creating a burned area map could prove to make a more accurate map in terms of area burned represented.

Burned-Area Mapping Using Post-Fire PlanetScope Images and a Convolutional Neural Network

Forest fires result in significant damage, including the loss of critical ecosystems and individuals that depend on forests. Remote sensing provides efficient and reliable information for forest fire detection on various scales. The purposes of this study were to produce burned-area maps and to identify the applicability of transfer learning. We produced a burned-area (BA) maps using single post-fire PlanetScope images and a deep learning (DL)-based algorithm for three cases in the Republic of Korea and Greece. Publicly accessible Copernicus Emergency Management Service and land cover maps were used as reference data for classification and validation. The DL model was trained using six schemes, including three vegetation indicators, and the data were split into training, evaluation, and validation sets based on a specified ratio. In addition, the model was applied to another site and assessed for transferability. The performance of the model was assessed using its overall accuracy. The U-Net model used in this study produced an F1-score of 0.964–0.965 and an intersection-over-union score of 0.938–0.942 for BAs. When compared with other satellite images, unburned and non-forested areas were accurately identified using PlanetScope imagery with a spatial resolution of approximately 3 m. The structure and seasonality of the vegetation in each target area were also more accurately reflected because of the higher resolution, potentially lowering the transferability. These results indicate the possibility of efficiently identifying Bas using a method based on DL with single satellite images.

An Automated Cropland Burned-Area Detection Algorithm Based on Landsat Time Series Coupled with Optimized Outliers and Thresholds

Given the increasingly severe global fires, the accurate detection of small and fragmented cropland fires has been a significant challenge. The use of medium-resolution satellite data can enhance detection accuracy; however, key challenges in this approach include accurately capturing the annual and interannual variations of burning characteristics and identifying outliers within the time series of these changes. In this study, we focus on a typical crop-straw burning area in Henan Province, located on the North China Plain. We develop an automated burned-area detection algorithm based on near-infrared and short-wave infrared data from Landsat 5 imagery. Our method integrates time-series outlier analysis using filtering and automatic iterative algorithms to determine the optimal threshold for detecting burned areas. Our results demonstrate the effectiveness of using preceding time-series and seasonal time-series analysis to differentiate fire-related changes from seasonal and non-seasonal influences on vegetation. Optimal threshold validation results reveal that the automatic threshold method is efficient and feasible with an overall accuracy exceeding 93%. The resulting burned-area map achieves a total accuracy of 93.25%, far surpassing the 76.5% detection accuracy of the MCD64A1 fire product, thereby highlighting the efficacy of our algorithm. In conclusion, our algorithm is suitable for detecting burned areas in large-scale farmland settings and provides valuable information for the development of future detection algorithms.

Sentinel 2 based burn severity mapping and assessing post-fire impacts on forests and buildings in the Mizoram, a north-eastern Himalayan region

The Increasing frequency and severity of forest fires worldwide highlights the need for more effective Burnt area mapping. Finding the effects of fire on vegetation and putting mitigation methods in place, depends on post-fire evaluation. In this study, the location of the burned regions and the severity of the fire were determined using high-resolution multi-spectral images from Sentinel 2 on Google Earth Engine (GEE) platform. Three widely used fire severity indices—differenced Normalized Burn Ratio (dNBR), Relativized Burn Ratio (RBR), and Relativized dNBR (RdNBR)—based on pre-fire Normalized Burn Ratio (NBR) and post-fire NBR—were computed and compared based on their accuracy using very high-resolution planet imagery fire points and equal number of random non fire points. Maps also validated with active fires, ground based photos and crowdsourced images. The accuracy (AUC) of the RdNBR map was 85%, RBR - 84% and dNBR −82%. The RdNBR index demonstrated highest level of accuracy. Then the loss to vegetation using pre-fire and post-fire NDVI was analysed. The analysis of pre-fire and post-fire NDVI provided insights into the extent of vegetation loss. The analysis of vegetation loss offered valuable information regarding the impact of fire on the affected areas. Google building dataset was used to monitor the percent of buildings under threat due to these fires. Around 8.77% of buildings were found in high severity region. Accurate mapping aids post-fire evaluation, guided mitigation strategies, and enhanced forest management and ecological restoration.

Geospatial assessment of forest fire impacts utilizing high-resolution KazEOSat-1 satellite data

Forest fires or wildfires frequently occur in Kazakhstan, especially in the months from June to September, damaging the forest resources. Burnt area mapping is important for fire managers to take appropriate mitigation steps and carry out restoration activities after the fire event. In this study, KazEOSat-1 high-resolution satellite datasets are used to map the burnt area in the regions of Kazakhstan. KazEOSat-1 satellite is in a Sun-synchronous orbit, consisting of four bands, namely blue, green, red, and NIR multispectral bands, in 4 m spatial resolution, while panchromatic data are in 1 m spatial resolution. This study examined three spectral indices, namely AVI, BAI, and GEMI, for mapping the burnt area based on the four spectral bands NIR, blue, red, and green of the KazEOSat-1 satellite datasets. The DN values for each band are used to determine TOA reflectance, which is then used as a basis for deriving the aforementioned spectral indices. The results of spectral indices, AVI, BAI, and GEMI are compared based on a discriminative index (M) for quantifying the effectiveness of each index based on burned area derived from KazEOSat-1 datasets. The spectral index BAI shows higher M values than other indices; therefore, the index BAI has the higher capability to extract the burned area as compared with AVI and GEMI. Accuracy was calculated based on the number of forest fire incidents that fell in burned and unburned areas, and the results indicate that BAI shows the highest accuracy, whereas AVI shows the lowest accuracy among them. Therefore, the BAI has the highest ability for extracting the burned area using the KazEOSat-1 satellite datasets. As the revisit time period of KazEOSat is 3 days, this study will be useful to map the burnt area and fire progression in Kazakhstan.

Mapping burned areas in Thailand using Sentinel-2 imagery and OBIA techniques

Monitoring burned areas in Thailand and other tropical countries during the post-harvest season is becoming increasingly important. High-resolution remote sensing data from Sentinel-2 satellites, which have a short revisit time, is ideal for accurately and efficiently mapping burned regions. However, automating the mapping of agriculture residual on a national scale is challenging due to the volume of information and level of detail involved. In this study, a Sentinel-2A Level-1C Multispectral Instrument image (MSI) from February 27, 2018 was combined with object-based image analysis (OBIA) algorithms to identify burned areas in Mae Chaem, Chom Thong, Hod, Mae Sariang, and Mae La Noi Districts in Chiang Mai, Thailand. OBIA techniques were used to classify forest, agricultural, water bodies, newly burned, and old burned regions. The segmentation scale parameter value of 50 was obtained using only the original Sentinel-2A band in red, green, blue, near infrared (NIR), and Normalized Difference Vegetation Index (NDVI). The accuracy of the produced maps was assessed using an existing burned area dataset, and the burned area identified through OBIA was found to be 85.2% accurate compared to 500 random burned points from the dataset. These results suggest that the combination of OBIA and Sentinel-2A with a 10 m spatial resolution is very effective and promising for the process of burned area mapping.

A novel deep Siamese framework for burned area mapping Leveraging mixture of experts

Due to the complexity of the areas and the diversity of the objects, traditional Burned Area Mapping (BAM) methods cannot provide promising results. Moreover, these methods focus on additional processing to improve their results, which is time-consuming and complex. Therefore, an advanced framework is needed to achieve accurate results in burned area mapping. In this context, this study proposes a novel Siamese-based mixture of expert networks for burned area mapping, which takes advantage of a mixture of experts (MoE) and position and channel attention mechanisms for deep feature generation. The proposed framework has two deep feature extractor channels to account for the bi-temporal multispectral pre- and post-fire input datasets. To evaluate the performance of the SMoE model, three multispectral Sentinel-2 were used in different countries. The results were compared with other advanced machine and deep learning models, including Light Gradient Boosting Machine (LGBM), Extreme Gradient Boosting (XGBoost), 2D-Siamse Convolutional Neural Network (2D-SCNN), and 3D-Siamse Convolutional Neural Network (3D-SCNN). The result of the burned mapping shows that the proposed model has a high effectiveness compared to other models, as it provides an average accuracy of more than 98% and 0.91 b y overall accuracy (OA) and kappa coefficient (KC) indices, respectively.

Regional-Scale Assessment of Burn Scar Mapping in Southwestern Amazonia Using Burned Area Products and CBERS/WFI Data Cubes

Fires are one of the main sources of disturbance in fire-sensitive ecosystems such as the Amazon. Any attempt to characterize their impacts and establish actions aimed at combating these events presupposes the correct identification of the affected areas. However, accurate mapping of burned areas in humid tropical forest regions remains a challenging task. In this paper, we evaluate the performance of four operational BA products (MCD64A1, Fire_cci, GABAM and MapBiomas Fogo) on a regional scale in the southwestern Amazon and propose a new approach to BA mapping using fraction images extracted from data cubes of the Brazilian orbital sensors CBERS-4/WFI and CBERS-4A/WFI. The methodology for detecting burned areas consisted of applying the Linear Spectral Mixture Model to the images from the CBERS-4/WFI and CBERS-4A/WFI data cubes to generate shadow fraction images, which were then segmented and classified using the ISOSEG non-supervised algorithm. Regression and similarity analyses based on regular grid cells were carried out to compare the BA mappings. The results showed large discrepancies between the mappings in terms of total area burned, land use and land cover affected (forest and non-forest) and spatial location of the burned area. The global products MCD64A1, GABAM and Fire_cci tended to underestimate the area burned in the region, with Fire_cci underestimating BA by 88%, while the regional product MapBiomas Fogo was the closest to the reference, underestimating by only 7%. The burned area estimated by the method proposed in this work (337.5 km2) was 12% higher than the reference and showed a small difference in relation to the MapBiomas Fogo product (18% more BA). These differences can be explained by the different datasets and methods used to detect burned areas. The adoption of global products in regional studies can be critical in underestimating the total area burned in sensitive regions. Our study highlights the need to develop approaches aimed at improving the accuracy of current global products, and the development of regional burned area products may be more suitable for this purpose. Our proposed approach based on WFI data cubes has shown high potential for generating more accurate regional burned area maps, which can refine BA estimates in the Amazon.

FLOGA: A Machine-Learning-Ready Dataset, a Benchmark, and a Novel Deep Learning Model for Burnt Area Mapping With Sentinel-2

FLOGA: A Machine-Learning-Ready Dataset, a Benchmark, and a Novel Deep Learning Model for Burnt Area Mapping With Sentinel-2

Burned area mapping in Different Data Products for the Southwest of the Brazilian Amazon

Fires affect the Amazon rainforest and cause various socio-environmental problems. Analyses of forest fire dynamics supporting actions to combat and prevent forest fires. However, many studies have reported discrepancies in the quantification of fire, especially in the tropics. We evaluated four operational products for estimating burned areas (MAPBIOMAS, MCD64A1, GABAM, and GWIS) in a part of the southwestern Brazilian Amazon. We used the year 2019 as a reference to assess the relative performance of each product through stratification by forest and non-forest areas. Statistical (Kolmogorov–Smirnov test) and geospatial analyses were performed using fuzzy similarity analysis and mapping of burned areas for forest and non-forest classes. The four products showed a divergence of up to 90.6% in the total area burned. MAPBIOMAS was the product with the largest area burned (3379 km²), and MCD64A1 detected the smallest area (325 km²). MAPBIOMAS and GABAM generally overestimates burn scars in forest areas compared to MCD64A1 and GWIS. Factors that influence the mapping of burned areas include cloud shadow, the spatial resolution of sensors, and external noises (drought and decomposition of bamboo forests). We highlight the importance of field validation when mapping imagery to differentiate the truly burned areas from targets with similar spectral behavior.

Validating MCD64A1 and FireCCI51 burned area mapping in the Qinghai-Tibetan Plateau

As an alpine area, surface fires significantly affect the fragile ecosystems of the Qinghai-Tibetan Plateau (QTP). However, no studies on the accuracy of any remote sensing burned area (BA) products on the QTP has not been evaluated, resulting in higher uncertainties in BA mapping. To address this, we used 56-view medium-resolution Landsat images of the QTP from 2001 to 2020 to verify the MCD64A1 and FireCCI51 products by stratified sampling. The results showed that MCD64A1 was more accurate than FireCCI51 for identifying small-scale (<2 km2) BAs within the QTP, whereas FireCCI51 was more accurate for identifying medium-scale (2–5 km2) and large-scale (>5 km2) BAs. Additionally, MCD64A1 was more accurate than FireCCI51 for identifying BAs in forests, FireCCI51 was more accurate than MCD64A1 for identifying BAs in grasslands, shrublands, and wetlands. These results can aid fire risk assessment in the region to improve ecosystem stewardship and inform further improvements in BA products.

Google Earth Engine Framework for Satellite Data-Driven Wildfire Monitoring in Ukraine

Wildfires cause extensive damage, but their rapid detection and cause assessment remains challenging. Existing methods utilize satellite data to map burned areas and meteorological data to model fire risk, but there are no information technologies to determine fire causes. It is crucially important in Ukraine to assess the losses caused by the military actions. This study proposes an integrated methodology and a novel framework integrating burned area mapping from Sentinel-2 data and fire risk modeling using the Fire Potential Index (FPI) in Google Earth Engine. The methodology enables efficient national-scale burned area detection and automated identification of anthropogenic fires in regions with low fire risk. Implemented over Ukraine, 104.229 ha were mapped as burned during July 2022, with fires inconsistently corresponding to high FPI risk, indicating predominantly anthropogenic causes.

Improving wildland fire spread prediction using deep U-Nets

Forest fires are able to cause significant damage to humans and the earth's fauna and flora. If a fire is not detected and extinguished before it spreads, it can have disastrous results. In addition to satellite images, recent studies have shown that exploring both weather and topography characteristics is crucial for effectively predicting the propagation of wildfires. In this paper, we present FU-NetCastV2, a deep learning convolutional neural network for fire spread and burned area mapping. This algorithm predicts which areas around wildfires are at high risk of future spread. With an accuracy of 94.6% and an AUC of 97.7%, the model surpassed the literature by 3.7% and exhibited a 1.9% improvement over our previous model. The proposed approach was implemented using consecutive forest wildfire perimeters, satellite images, Digital Elevation Model maps, aspect, slope and weather data.

FOREST FIRE BURNT AREA EXTRACTION USING FUZZY INTEGRATION OF MULTI-SENSOR SATELLITE DATA FOR THE HIMALAYAN STATE

Abstract. Burnt area assessment due to forest fires is an important aspect to estimate the extent of loss of biodiversity which has become feasible even in hilly and inaccessible areas with the help of geospatial technologies. But satellite data also has some limitations as it increases commission error by misclassifying non-burnt areas as burnt areas. To reduce this commission error, present study has attempted to integrate multi-sensor satellite data to characterize and extract forest fire burnt areas in Uttarakhand which is a fire prone hilly state in Western Himalaya. Landsat-8 and Sentinel-2 optical datasets have been used to calculate eleven vegetation/burn indices to identify burn patches for fire season of 2022 (February to June). These vegetation/burn indices have been calculated from Landsat-8 and Sentinel-2 datasets and integrated using Fuzzy Logic Modelling to get characterized forest fire burnt area maps. Accuracy assessment has been done using Moderate Resolution Imaging Spectroradiometer (MODIS) and Visible Infrared Imaging Radiometer Suite (VIIRS) active fire points for the characterized map of burnt area by Landsat-8, Sentinel-2 and combining indices from both sensors. The fuzzy map of burnt area using Landsat-8 showed the accuracy of 66.25%, while Sentinel-2 showed accuracy of 59.79% and the integration of fuzzy burnt area maps of both sensors showed the highest accuracy of 79.66%. This information of characterized burnt areas of a region can help forest managers to identify high vulnerable areas to focus on during the fire season to prevent the losses to natural resources, life and property in the region.

Total-variation regularized U-Net for wildfire burned area mapping based on Sentinel-1 C-Band SAR backscattering data

Previous studies have shown that Synthetic Aperture Radar (SAR) is able to detect burned areas, serving as a key data source for monitoring active wildfires in situations where optical sensors are hindered by dense smoke or cloud cover. Radar remote sensing provides rich and useful data on historical burned areas, which are critical for large-scale wildfire burned area mapping. This study aims to unveil the potential inherent in Sentinel-1 C-Band SAR data and to investigate the impact of reference masks on wildfire burned area mapping. SAR images frequently exhibit disruptive speckle noise, which manifests as isolated pixels or small regions, thereby hindering the seamless identification of burned areas through SAR data. To mitigate noise and enhance the connectivity of SAR-based burned area delineation, we propose a novel approach: a Total-Variation (TV) regularized U-Net model, tailored to learn from noisy pseudo masks derived from Sentinel-1 C-Band SAR backscattering data. To validate its efficacy, we assembled a dataset encompassing 16 geographically diverse wildfire events, incorporating SAR-based kMap, a SAR-based binary label (designated as SARREF), and an optical-based binary label (referred to as OptREF). The kMap, serving as a relative indicator of SAR backscattering change, is defined as the divergence between post-fire SAR backscattering and the temporal average, normalized by the temporal standard deviation of pre-fire SAR time series. Utilizing SAR-based kMap as input and SARREF or OptREF as labels, we systematically explored the impact of TV regularization on SAR-based burned area mapping. Our experimental findings highlight several key insights: 1) Vanilla U-Net considerably elevates the F1 score from 0.61 to 0.66 when compared to threshold-based approaches; 2) The TV-regularized U-Net exhibits a notable capacity for enhancing both accuracy and the connectivity of SAR-based burned areas. When trained with noisy SARREF, it achieves a substantial F1 score of 0.69 and an IoU score of 0.53. 3) In scenarios involving OptREF, the vanilla U-Net attains peak performance with the highest F1 (0.75) and IoU (0.61) scores, with the expected outcome of TV regularization yielding no additional enhancement.

CYCLE-GAN BASED FEATURE TRANSLATION FOR OPTICAL-SAR DATA IN BURNED AREA MAPPING

Abstract. For the management of the forest and the assessment of impacts on ecosystems, post-fire burned area mapping is crucial for sustainable environment and forestry. While optical remote sensing data has been extensively used for monitoring forest fires due to its spatial and temporal resolutions, it is susceptible to limitations imposed by poor weather conditions. To overcome this challenge, the complementary use of optical and Synthetic Aperture Radar (SAR) data is beneficial, as SAR can penetrate clouds and capture images in all-weather conditions. However, SAR lacks the necessary spectral features for comprehensive forest fire monitoring and burned area mapping. To overcome these limitations, this study proposes a Cycle-Consistent Generative Adversarial Networks (Cycle-GAN) based deep feature translation method for burned area mapping by combining optical and SAR data. This approach allows for the retrieval of precise information of interest with a level of precision that cannot be achieved by either optical or SAR data alone. The Cycle-GAN uses a cyclic structure to transfer data from one domain (optical) to another domain (SAR) into the same feature space. As a result, it can maintain its spectral characteristics while providing ongoing and current information for monitoring forest fires. For this purpose, Burn Area Index (BAI), Mid Infrared Burn Index (MIRBI), Normalised Burn Ratio (NBR) were determined using optical data and image translation was performed with Cycle-GAN on SAR data. The accuracy of the fake BAI, MIRBI and NBR spectral burn indices determined from the SAR was established by correlating the original spectral burn indices determined from the optical data. The results demonstrate a significant correlation between the real and generated fake burn indices, particularly with a noteworthy correlation coefficient of 0.93 observed for the NBR index. In addition, the findings validate the effectiveness of the generated indices in accurately representing and quantifying the extent of burned areas.