Open AccessArticle

Mapping Burned Area in the Caatinga Biome: Employing Deep Learning Techniques

Washington J. S. Franca Rocha

^1,*

Rodrigo N. Vasconcelos

^1,2

Soltan Galano Duverger

^2,3

Diego P. Costa

^1,2,4

Nerivaldo A. Santos

^1,2

Rafael O. Franca Rocha

^1,2

Mariana M. M. de Santana

⁵

Ane A. C. Alencar

⁶

Vera L. S. Arruda

⁶,

Wallace Vieira da Silva

⁶,

Jefferson Ferreira-Ferreira

⁷

Mariana Oliveira

⁷

Leonardo da Silva Barbosa

⁷

and

Carlos Leandro Cordeiro

⁷

Postgraduate Program in Earth Modeling and Environmental Sciences—PPGM, State University of Feira de Santana—UEFS, Feira de Santana 44036-900, Brazil

GEODATIN–Data Intelligence and Geoinformation, Bahia Technological Park Rua Mundo, 121–Trobogy, Salvador 41301-110, Brazil

Multidisciplinary and Multi-Institutional Postgraduate Program in Knowledge Diffusion (DMMDC/UFBA), Federal University of Bahia—UFBA, Salvador 40110-100, Brazil

⁴

Interdisciplinary Center for Energy and Environment (CIEnAm), Federal University of Bahia UFBA, Salvador 40170-115, Brazil

⁵

Forest Engineering Institute (FEI/UEAP), State University of Amapá—UEAP, Av. Pres. Getúlio Vargas, 650 Centro, Macapá 68900-070, Brazil

⁶

Instituto de Pesquisa Ambiental da Amazônia (IPAM), SCN 211, Bloco B, Sala 201, Brasília 70836-520, Brazil

⁷

World Resources Institute Brasil, Rua Cláudio Soares, 72 Cj. 1510, São Paulo 05422-030, Brazil

Author to whom correspondence should be addressed.

Fire 2024, 7(12), 437; https://doi.org/10.3390/fire7120437

Submission received: 14 October 2024 / Revised: 12 November 2024 / Accepted: 22 November 2024 / Published: 27 November 2024

(This article belongs to the Special Issue Machine Learning (ML) and Deep Learning (DL) Applications in Wildfire Science: Principles, Progress and Prospects)

Download

Browse Figures

Figure 1
Map of the boundaries of the Caatinga biome. "> Figure 2
Overview of the method for classifying burned areas in Caatinga. "> Figure 3
The Multi-Layer Perceptron Network‘s structure involves using the spectral bands (RED, NIR, SWIR1, and SWIR2) as input layers and the classes burned and unburned as the output layers. "> Figure 4
The Multi-Layer Perceptron Network‘s structure involves using the spectral bands (RED, NIR, SWIR1, and SWIR2) as input layers and the classes burned and unburned as the output layers. (A) depicts the cumulative burn area from 1985 to 2023. (B) in contrast, showcases the annual burn area over the same temporal range. "> Figure 5
The annual distribution of annual burned class areas in the Caatinga biome from 1985 to 2023. "> Figure 6
The annual distribution of burned areas by land use and land cover types in the Caatinga biome from 1985 to 2023. "> Figure 7
The paper presents the spatial distribution of fire frequency in Brazil from 1985 to 2023, including the corresponding burned area and proportion by frequency class. (A) shows the map of fire frequency throughout the Caatinga biome, while (B) presents the classes of fire frequency by area and their corresponding percentages. "> Figure 8
The figures depict the spatial association between accumulated burn scars and various climate parameters. (A) illustrates the correlation between burn scars and accumulated precipitation. (B) showcases the relationship between accumulated burn scars and climate water deficit. Lastly, (C) presents the correlation between burn scars and reference evapotranspiration. ">

Versions Notes

Abstract

The semi-arid Caatinga biome is particularly susceptible to fire dynamics. Periodic droughts amplify fire risks, while anthropogenic activities such as agriculture, pasture expansion, and land-clearing significantly contribute to the prevalence of fires. This research aims to evaluate the effectiveness of a fire detection model and analyze the spatial and temporal patterns of burned areas, providing essential insights for fire management and prevention strategies. Utilizing deep neural network (DNN) models, we mapped burned areas across the Caatinga biome from 1985 to 2023, based on Landsat-derived annual quality mosaics and minimum NBR values. Over the 38-year period, the model classified 10.9 Mha (12.7% of the Caatinga) as burned, with an average annual burned area of approximately 0.5 Mha (0.56%). The peak burned area reached 0.89 Mha in 2021. Fire scars varied significantly, ranging from 0.18 Mha in 1985 to substantial fluctuations in subsequent years. The most affected vegetation type was savanna, with 9.8 Mha burned, while forests experienced only 0.28 Mha of burning. October emerged as the month with the highest fire activity, accounting for 7266 hectares. These findings underscore the complex interplay of climatic and anthropogenic factors, highlighting the urgent need for effective fire management strategies.

Keywords:

forest fire; wildfire; semiarid; drylands; burned area; fire regime; Caatinga vegetation

1. Introduction

The global frequency and severity of wildfires are on the rise [1,2,3], leading to significant impacts on ecosystem services, the climate, the economy [4] and human health [5,6]. These fires are mainly caused by human activity in extreme weather conditions, characterized by high temperatures, low rainfall, and abundant fuel availability, all of which contribute to the rapid spread of these destructive blazes [1,7,8]. Human-related ignition sources, such as land-use changes and proximity to agricultural, residential, and infrastructure areas, are responsible for a majority of large, catastrophic fires [9,10,11]. It is imperative to address the complex human factors driving wildfires to effectively mitigate their detrimental environmental, social, and economic effects.

In South America, Brazil exhibits the highest frequency of fires [12,13], a phenomenon that becomes even more concerning with the weakening of environmental policies and rising deforestation rates. These factors are likely to exacerbate the occurrence of wildfires throughout the country [14]. The disturbances related to the increase in fires are especially concentrated in biomes such as the Cerrado, the Amazon, and the Pantanal [15,16]. While the Cerrado and Pantanal are recognized as biomes adapted to fire [17,18,19], the intensification of fire regimes and human pressure may compromise their resilience.

It is important to note that the Amazon, despite being a vital ecosystem, is particularly sensitive and has been extensively researched, highlighting the need to understand the effects of fires on its ecological balance. Research often focuses on the vulnerability of these environments in the face of increasing fire incidences, particularly in the tropical regions of the Amazon, where the occurrence of extreme fires is expected to rise [20,21,22]. However, the implications of increased fire frequency and changes in fire patterns within highly seasonal environments, such as the drylands in the Caatinga biome of Brazil, are not yet fully comprehended [23,24,25]. This is primarily because there is a lack of focused efforts aimed at systematically mapping and monitoring burned areas in these regions [16,26,27,28].

The Caatinga biome, in northeastern Brazil, renowned for its distinct and diverse ecosystems, is particularly vulnerable to wildfires [29,30], a fact that is often exacerbated by human activities [31,32]. Being one of the world’s most imperiled ecosystems, the Caatinga biome is especially at risk of wildfires due to widespread deforestation, as well as intensive agricultural and livestock activities [29,33]. This highlights the pressing need for improved fire management and environmental conservation measures in Brazil. Therefore, it is imperative to gain a comprehensive understanding of the spatial and temporal dynamics of fire in the Caatinga [34] to support effective fire management and conservation strategies in this ecologically vital region.

To effectively mitigate the impacts of fires, it is essential to understand the historical patterns of fire occurrence, along with their driving factors and consequences. In Brazil, it is important to recognize that local communities often initiate fires for land clearance for agricultural purposes; however, these fires can quickly escalate out of control, leading to significant environmental degradation and loss of biodiversity. Therefore, closely monitoring the burned areas in the Caatinga is crucial for gaining a comprehensive understanding of the underlying causes of these fires, particularly the role of human activities. This knowledge will be vital for developing effective mitigation strategies that address both environmental protection and community needs.

Remote sensing techniques have become an essential tool for monitoring and managing wildfires, providing valuable data on the spatial and temporal dynamics of burned areas [35]. In recent years, machine learning and deep learning approaches have shown great promise in modeling and predicting wildfire behavior, particularly in regions with complex environmental and anthropogenic factors [36,37,38,39]. Deep neural networks, along with other deep learning methodologies, are recognized for their sophisticated approach inspired by the structure and function of the human brain [40]. These models comprise interconnected layers of artificial neurons that process input data to extract increasingly abstract and high-level features. The key strength of deep neural networks lies in their ability to autonomously learn and extract necessary representations from raw data, enabling them to perform tasks such as feature detection and classification effectively [41].

In remote sensing, deep neural networks have demonstrated significant potential for various applications, including mapping burn scars [42,43,44,45]. These networks can analyze large volumes of satellite imagery to identify and classify areas affected by fires accurately. By harnessing the deep learning capabilities of deep neural networks, researchers can efficiently analyze temporal and spatial patterns of burned areas, thereby facilitating improved monitoring and management of fire-affected regions. Deep neural networks’ automated feature extraction and classification capabilities notably enhance the precision and speed of burn scar mapping, providing valuable insights for environmental management and disaster response efforts.

This research represents a significant milestone, as it is the first effort to map burned areas within the Caatinga biome over a span of 38 years. This study is part of the collaborative MapBiomas Brasil project, specifically focusing on the Caatinga chapter. MapBiomas is a consortium of public and private institutions comprising remote sensing experts with the main goal of understanding the dynamics of land use in Brazil through the use of cloud computing and data science. The research makes use of cloud processing algorithms hosted on Google Earth Engine and Google Cloud Platform [46].

Our research endeavors to produce maps that offer vital insights into ecosystem changes, fire risk assessment, fire management guidance, and policy-informing data regarding fire patterns in the Caatinga biome of Brazil. We plan to utilize a deep learning-based approach to map monthly burned areas by employing satellite imagery to differentiate between burned and unburned regions. The key objectives of this research include assessing the effectiveness of a fire detection model, examining the spatial and temporal patterns of burned areas, providing valuable input for efficient fire management and prevention strategies, and making the developed model and related data accessible to the public for further research and applications.

2. Materials and Methods

2.1. Study Area

The Caatinga biome covers an area of approximately 86.3 million hectares, accounting for about 10% of Brazil’s total land area. It is situated in the northeastern region of the country, extending from 3° to 17° S in latitude and from 35° to 45° W in longitude (Figure 1). Geographically, the Caatinga is bordered to the east by “Tropical Moist Broadleaf Forests” [47], corresponding to the Atlantic Forest biome. To the west, it transitions into the Cerrado biome, marked by “Tropical Grasslands, Savannas, and Shrublands”, though in the northwest, it also contains a portion of “Tropical Moist Broadleaf Forests” [47]. This ecological gradient emphasizes the transitional nature of the Caatinga, which lies between more humid eastern ecosystems and the drier savannas and shrublands to the west [48].

The region is characterized by sporadic rainfall patterns and prolonged arid periods, which are key climatic factors that shape its unique ecosystems [49]. The climate in this region significantly differs from the typical zonal climates usually found at similar latitudes. It is characterized by high temperatures, significant evapotranspiration, and unpredictable precipitation patterns ranging from 240 mm to 1500 mm annually, with most rainfall occurring within two or three consecutive months [50]. These climatic patterns lead to severe droughts and a landscape marked by an intricate network of intermittent rivers.

The term “Caatinga”, which means ‘whitish forest’ in the indigenous Tupi language, reflects the deciduous nature of most trees and shrubs during the dry season [51]. The landscape is dominated by shrubby, branched, and thorny vegetation, often adorned with bromeliads, euphorbias, and cacti. Taxonomically, the predominant vegetation cover type within Caatinga biome is classified as “Xeric Shrublands” [47] which is also referred to as “Steppe Savanna” [52]. This classification highlights the semi-arid climate and the drought-adapted nature of the Caatinga flora. Additionally, this unique biome also hosts the largest area of “Tropical Dry Broadleaf Forests” in the Americas, underscoring its ecological significance [53,54,55].

The Caatinga biome, renowned for its rich biodiversity and high levels of endemism. This unique biome harbors a diverse range of species, with 3150 plants, 276 ants, 386 fishes, 98 amphibians, 79 reptiles, 548 birds, and 183 mammals [53]. However, the Caatinga represents one of Brazil’s most threatened biomes due to centuries of unsustainable land and resource management, like illegal logging, grazing, and conversion to agricultural and pasture lands. In response to the rapid deforestation, the Ministry of the Environment has set up federal and state Conservation Units and encouraged sustainable alternatives to preserve the biodiversity of the biome. Yet, despite these endeavors, the Caatinga remains one of Brazil’s least protected biomes, with just over 1% of its area designated as Integral Protection Units, and many existing Conservation Units, especially the Environmental Protection Areas, need to be more effectively enforced.

Although recent efforts by both governmental and non-governmental organizations have started to tackle the urgent conservation needs of this biome, there still needs to be a more scientific understanding of this distinct and delicate ecosystem. The Caatinga remains one of the least studied ecosystems in Latin America and among the most neglected within the country [56]. The biome’s inherent vulnerability, marked by shallow soils, periodic droughts, high evapotranspiration, and erratic rainfall, is compounded by the absence of technical support for rural land management.

2.2. General Strategy

Using all available Landsat 5, 7, 8, and 9 imageries, we applied deep neural network (DNN) models to detect and map burned areas across the Caatinga biome from January 1985 to December 2023. These DNN models utilize advanced artificial intelligence and machine learning algorithms to perform deep learning-based classifications, allowing for enhanced accuracy in capturing complex patterns, such as burned area detection. The deep learning approach significantly improves the performance and precision of burned area mapping, especially when dealing with long-term and large-scale datasets.

The methodology for mapping monthly burned areas across the Caatinga followed a six-step process. Initially, classification regions were defined by biome. Annual high-quality Landsat mosaics were then constructed, preprocessed using Google Earth Engine. Training samples of spectral signatures for both burned and unburned areas were collected from the mosaics, with results exported to a Google Cloud Storage Bucket. A deep neural network (DNN) prediction model was subsequently trained and developed. Post-classification routines, including the application of masks and spatial filters, were used to refine the classification outputs. Finally, an accuracy assessment was conducted to evaluate model performance. The trained DNN models were applied to process burn scar mapping and produce a comprehensive dataset of monthly burned areas across Brazil from 1985 to 2023. The methodological framework for detecting and mapping burned areas within the Caatinga biome is illustrated in Figure 2.

2.3. Classification and Sample Regions

The analysis of climate, land cover, and land use impacts on fire patterns, as well as the spectral properties of burned areas, was conducted using an integrated classification process based on hydrographic basins. Specifically, level 2 basins, as defined in the National Water Resources Plan, were used with data provided by Brazil’s National Water and Basic Sanitation Agency—ANA, free available online (https://metadados.snirh.gov.br) for downloaded (accessed on 19 September 2024). These basins were selected as training units for the model to classify burned areas, as they offer enhanced accuracy by grouping regions with environmental similarities such as vegetation cover, soil composition, and climate patterns. This delineation of regions during the pre-processing stage improves the precision of burned area classification and enhances the understanding of fire dynamics in ecologically comparable regions.

Surface reflectance Landsat mosaics spanning from 1985 to 2023 were utilized for classification analysis, focusing specifically on scenes covering the Caatinga biome. This method was enabled by utilizing the Google Earth Engine platform (GEE; https://earthengine.google.com, accessed on 19 September 2024), which leverages Google‘s cloud infrastructure and a JavaScript-based language to access and analyze extensive global geospatial datasets [46]. To mitigate issues related to cloud cover, the Landsat Quality Assessment band was employed to generate annual quality mosaics, effectively masking problematic pixels. For each location, the pixel exhibiting the minimum Normalized Burn Ratio (NBR) value was selected, as the NBR index is well established for its efficacy in detecting fire scars and assessing burn severity [57,58]. These annual quality mosaics subsequently served as the foundation for training and applying deep neural network (DNN) models, facilitating the mapping of burned areas across the Caatinga biome.

The sampling strategy incorporated global burned area products from NASA‘s Moderate Resolution Imaging Spectroradiometer (MODIS), specifically the MCD64A1 product, from NASA (https://ladsweb.modaps.eosdis.nasa.gov/missions-and-measurements/products/MCD64A1 accessed on 19 September 2024), as a reference for burned areas between 2000 and 2020. Additionally, active fire data from the National Institute for Space Research (INPE) in Brazil (https://terrabrasilis.dpi.inpe.br/queimadas accessed on 19 September 2024), which provides daily fire activity records, were included. Prior to 2000, fire reference data were not utilized due to insufficient spatial confidence in the available datasets. Landsat scene mosaics with significant burned areas and active fires were selected to establish a spectral library for training the DNN models. Samples were stratified by Landsat sensor to accurately capture the unique spectral characteristics of each sensor, resulting in a final spectral library divided into stacks that served as input for the classification process.

A spatial filter was used to refine the burned area classification, removing small noise features and filling in minor gaps. After evaluating the classification results, post-classification routines were applied to reduce errors from land cover types with similar spectral signatures to recently burned areas, such as water, urban areas, and certain crops. Specific rules were defined per biome to remove pixels incorrectly classified as burned within distinct land cover and land use classes from MapBiomas Collection 8.0 (http://mapbiomas.org), dataset (accessed on 19 September 2024). This post-processing reduced the original burned area estimate. Figure 2 illustrates the workflow, including the definition of classification regions and the collection of burned and unburned area samples, for the application of post-classification filters to refine results based on land use and land cover classes.

2.4. Deep Learning Architecture Model

The classification model used was the deep neural network, a computational model capable of deep learning and visual pattern recognition. The Multi-Layer Perceptron Network is a deep neural network model that consists of interconnected layers of computational units, where each node in one layer is connected to a node in the next layer. A Multi-Layer Perceptron (MLP) is a type of deep neural network with a feedforward architecture.

The model has three main layers: input, hidden, and output layers, where each neuron in a layer is connected to every neuron in the next layer (Figure 3). The input layers used the RED, NIR, SWIR1, and SWIR2 spectral bands. This layer receives the raw data, which is then processed through the hidden layers using weighted sums, biases, and non-linear activation functions, allowing the MLP to capture complex, non-linear relationships. The output layer produces the final predictions; the layers classify the areas as burned or unburned. The network is trained using backpropagation to minimize the error between predicted and actual values.

The burned area mapping algorithm had two main steps: training and prediction. During the training phase, key parameters were defined based on prior testing, including learning rate, batch size, number of iterations, and classification inputs. The inputs used in the classification model were the surface reflectance spectral data from the annual quality mosaics, along with the training samples of burned and unburned areas.

We conducted our analysis using powerful computational resources, including graphics processing units and specialized hardware for parallel processing. The infrastructure used had 8 vCPUs, 32 GB of RAM, and 200 GB of additional disk space. We accessed GPUs through a virtual machine environment on the Google Cloud Platform, a suite of cloud computing services.

2.5. Data Analysis

Our post-classification processing, a meticulously precise methodology, was conducted to produce monthly and yearly burned area maps. This involved extracting date information for burned pixels from the annual quality mosaics, which were created using the minimum Normalized Burn Ratio value. The yearly fire scar maps represent the combination of all monthly burned areas yearly.

The cumulative burned area data, a comprehensive compilation, were meticulously organized by totaling the burned area for each year. Each pixel was counted only once as burned, regardless of any multiple fire events. These data were then organized by land cover and land use type based on the intersection of fire occurrences with the land cover and land use classes from MapBiomas Collection 8.0 for the most recent year of the study period.

The fire recurrence data were generated by consolidating annual burned area information into a single map class. Each class represents the number of times a pixel was burned during the study period. For example, class 1 indicates pixels burned once, class 2 indicates pixels burned twice, and so on. These classes provide a comprehensive view of the fire frequency and its impact on the landscape. Additionally, the fire frequency data were overlaid with the MapBiomas Collection 8.0 land cover and land use map for the year 2023, the final year of the data series, to examine fire patterns across different land types.

We perform analyses at the biome level: the extent of burned areas, the monthly fire season, the burn area in LULC, and fire frequency. We also calculate the accumulated burn area for the states and municipalities the Caatinga biome covers.

To evaluate the relationship between burned areas and climatic variables, we developed scripts in Google Earth Engine to calculate the Pearson correlation. We selected the variables precipitation accumulation, reference evapotranspiration, and climate water deficit for analysis. We utilized data from TerraClimate, a monthly climate and climatic water balance for global terrestrial surfaces system developed by the University of Idaho. This dataset employs climatically aided interpolation, combining high-spatial resolution climatological normals from the WorldClim dataset. We used the 30-year average time series to conduct the present analysis. These variables are essential for understanding the relationship between climate and fire behavior, as rainfall significantly influences vegetation growth cycles and biomass accumulation. Meanwhile, evapotranspiration and water deficit emphasize moisture levels and the vulnerability to fire. This approach offers a nuanced perspective on fire dynamics, enhancing adaptive fire management and contributing to broader strategies for environmental protection within the biome.

2.6. Validation Analysis

The assessment of burned area mapping was conducted within the Caatinga biome, focusing on a thoughtfully selected range of years from 2007 to 2023. These years were intentionally chosen to encompass periods reflecting the highest and lowest extents of burned areas and to incorporate data from diverse sensor sources. This extensive temporal range facilitated a robust and thorough fire scar mapping process evaluation. During the sample selection process, a separate set of validation samples was identified each year, following the rigorous criteria described earlier. This varied dataset was then used to carry out comprehensive and independent analyses for each year, ensuring the reliability and accuracy of the overall assessment (see more details in Supplementary Materials Table S1).

Evaluating fire scar mapping accuracy in the Caatinga biome involves using multiple key performance metrics, including overall accuracy, precision, recall, F1 score, and the Jaccard index. These metrics provide insights into various aspects of the mapping process, collectively contributing to the fire detection methods’ reliability and efficacy. Below, we will present a concise overview of the metrics in use.

Overall accuracy measures the percentage of instances correctly classified out of the total. It provides a broad evaluation of the mapping performance, indicating how well the classification model distinguishes between fire-affected and unaffected areas. A high overall accuracy suggests that most classifications are correct, making it a vital indicator of the model’s overall reliability.

Precision, which reflects the predictive value, measures the percentage of accurate positive fire detections out of all positive detections made by the model. This metric is crucial for evaluating the accuracy of fire scar mapping, as it indicates the probability that areas identified as fire scars have truly been impacted by fire. High precision reduces the occurrence of false positives, ensuring that the identified fire scars accurately represent actual fire events.

The recall metric, also called sensitivity, measures the percentage of actual fire-affected areas that the model correctly identifies. It indicates the model’s ability to capture all relevant instances of fire damage. Maintaining a high recall is essential to ensure that important fire events are noticed during mapping, providing a comprehensive understanding of fire impacts within the ecosystem.

The F1 score is a metric that combines precision and recall into a single value by using their harmonic mean. It is beneficial when both precision and recall are equally important, as it provides a comprehensive assessment of the model’s performance. A high F1 score indicates that the model effectively balances accurately detecting fire scars and minimizing false identifications, making it a reliable indicator for evaluating the model’s overall effectiveness.

The Jaccard index, or the intersection over union, is a metric used to evaluate the spatial correspondence between predicted and actual fire scar areas. This index is determined by dividing the intersection size of the expected and actual fire-affected regions by the size of their union. The Jaccard index provides a spatial assessment of the model’s performance, indicating the agreement between the identified fire scars and the ground truth. Higher Jaccard index values suggest that the model’s predictions closely align with the real-world fire-affected areas, making it a valuable measure for evaluating the spatial accuracy of the fire scar mapping process.

3. Results

3.1. General Patterns

Over the course of 38 years, a meticulously crafted Landsat-based dataset, generated using state-of-the-art deep learning classification for the Caatinga biome, was used to analyze the annual variability of burned area (Figure 4A).

The data, a testament to the precision of our research, revealed that 10.9 Mha, equivalent to 12.7% of the Caatinga, experienced burning at least once between 1985 and 2023. On average, 481,609 hectares per year, or 0.56% of the Caatinga, were affected by fire annually.

The data reveal a significant yearly variation in the burned area, ranging from about 183,602 hectares in 1985 to a peak of around 899,109 hectares in 2021. This wide range suggests that the occurrence and intensity of fires have fluctuated considerably, reflecting the intricate interplay of climatic, environmental, and human factors that impact fire activity. The annual burned area does not exhibit a simple linear trend but displays notable year-to-year variability, highlighting the unpredictable nature of fire incidents (Figure 4B).

Despite significant variability, there was a noticeable increase in burned areas over the study period. Lower burned areas were generally recorded in the early years of the time series, particularly in the late 1980s and early 1990s. However, from the late 1990s onwards, there has been an apparent increase, with several years showing exceptionally high burned areas. This trend suggests worsening fire conditions, possibly influenced by global climate change, which is known to exacerbate fire-prone conditions through rising temperatures and altered precipitation patterns (Figure 4B).

Several years have notably higher burned areas, such as 1986, 1987, 1992, 1993, 2007, 2015, 2021, and 2023. These peaks may be attributed to extreme weather events, prolonged drought, or increased human activities like land clearing and agricultural burning. Identifying the specific causes of these peaks is crucial for devising effective fire prevention and response strategies (Figure 4B).

Over the study period, the average annual burned area was approximately 481,609 hectares, with a standard deviation of about 194,210 hectares, indicating significant variability. The median burned area was 468,502 hectares, suggesting a relatively symmetrical distribution with some years experiencing exceptionally high fire activity. This statistical distribution emphasizes that while most years align with the average, extreme events significantly influence the data, representing periods of intense fire activity that could have profound ecological and socio-economic impacts.

The analysis has also pinpointed periods of decreased burned areas, notably in 1985, 1988, 1991, 2011, and 2013 (Figure 4B). These periods of reduced fire activity might be due to favorable climate conditions, effective fire management policies, or a mix of both. These results indicate that taking proactive measures to reduce fire impacts is achievable, underscoring the significance of ongoing investment in fire prevention and control efforts.

When examining different decades, it becomes apparent that recent years have seen significantly larger areas affected by wildfires compared to the 1980s and 1990s. This escalation in fire activity in recent times may be attributed to changes in land use, increased human presence in fire-prone regions, and the broader impacts of climate change. The upward trajectory of fire incidents emphasizes the immediate requirement for adaptable fire management strategies to effectively address the changing fire patterns influenced by natural and human factors.

3.2. Extent of Burned Areas

The results obtained in relation to the size of the burned areas over time revealed that the most prevalent class in terms of total area is “100–500 ha”, which encompasses approximately 3,439,598 hectares (3.4 Mha) (Figure 5). This substantial representation indicates that fires of medium size are common in the Caatinga biome. Following closely is the “10–50 ha” class, covering about 3,366,692 hectares, highlighting the significance of smaller fire scars in the fire regime of this biome. The “<10 ha” class comes next, with an area of approximately 3,253,119 hectares, emphasizing the prevalence of small fires.

The category “1000–5000 ha” is also noteworthy, covering a total area of about 3,086,614 (Figure 5). This indicates that, although less frequent than smaller fires, larger fire incidents still play a significant role in the fire landscape of the Caatinga. The categories “50–100 ha” and “500–1000 ha” encompass 1,510,295 hectares and 1,434,706 hectares, respectively, underscoring that moderate-sized fire scars are quite common and substantially contribute to the overall burned area.

The categories “10,000–50,000 ha” and “5000–10,000 ha” represent smaller areas than the more common categories, with 1,274,538 hectares and 845,953 hectares, respectively (Figure 5). This indicates that larger fires still occur, although less frequently, and can cover significant areas. The “50,000–100,000 ha” and “≥100,000 ha” categories have the smallest total mapped areas, with 232,045 hectares and 217,493 hectares, respectively. This suggests that significant fire events are uncommon within the Caatinga biome.

Overall, while there is no consistent linear trend, there are periods of increased fire activity, and large-scale fires (over 50,000 hectares) were relatively more common in the earlier part of the series compared to the last decade. This suggests a shift in the fire regime, with recent years still experiencing high fire activity but with a broader range of burn sizes.

3.3. Monthly Fire Season Patterns

Examining monthly fire scar areas in the Caatinga biome demonstrates clear temporal trends. The data indicate a substantial rise in fire scar areas during the spring and early summer. Notably, October shows the most significant area affected by fire scars, totaling 7,266,083 hectares. This is trailed by November, with a total area of 3,707,726 hectares, and September, 3,418,562 hectares (Table 1). Substantial fire activity in these months is likely influenced by the transition from the dry to the rainy season when accumulated dry vegetation provides ample fuel for fires.

Autumn and early winter exhibit significantly smaller fire scar areas. In particular, April has the lowest total area of fire scars at 51,178 hectares, followed by June with 77,631 hectares and May with 83,345 hectares (Table 1). These months align with the rainy season in the Caatinga, leading to decreased fire activity due to higher moisture levels in the vegetation and soil. The decrease in fire occurrence during this period emphasizes the strong correlation between precipitation patterns and fire dynamics in the biome.

Since July, there has been a consistent increase in the areas affected by fire scars, reaching a peak in the subsequent months. Notably, July alone witnessed a significant rise, with 240,352 hectares impacted by fire scars. This trend continues into August, with a substantial increase to 1,043,449 hectares. The gradual intensification in fire activity from July onwards is consistent with the drying conditions following the rainy season, heightening the susceptibility of vegetation to ignition and the spread of fires.

Most fire scars, in terms of total hectares affected, are concentrated in October, November, and September. These months combined account for most of the annual fire scar areas in the Caatinga biome. October alone represents the most significant area, highlighting its critical period for fire management interventions. The high fire activity during these months requires focused efforts to reduce fire risks, including controlled burns and community awareness programs to prevent accidental ignitions (Table 1).

3.4. Burn Area in LULC

The annual distribution of burned areas by land use and land cover types in the Caatinga biome from 1985 to 2023 is presented in Figure 6. This chart provides valuable insights into the dynamics of fire occurrence across different land cover classes over nearly four decades.

The analysis of fire scars across different land use and land cover types in the Caatinga biome from 1985 to 2023, based on the figure, shows that savanna vegetation was significantly impacted by fires, with a total of approximately 9.8 million hectares burned during the 38-year study period. This highlights the savanna’s susceptibility to large-scale and frequent fire events, with an annual average of around 258,000 hectares affected by fires, ranging from a minimum of 76,000 hectares to a maximum of 768,000 hectares in individual years.

Grasslands recorded a total of 535,473 hectares burned, averaging 13,730 hectares per year, with a range of 2424 to 45,842 hectares. While fire plays a natural role in maintaining the ecological balance of savannas and grasslands, promoting new growth and biodiversity, the high frequency of fire events can lead to significant biodiversity loss. However, excessive fires may disrupt the delicate balance of these ecosystems and threaten native species.

In contrast, forest vegetation experienced far smaller burned areas, with a total of approximately 288,479 hectares affected by fires throughout the period. On average, 7397 hectares of forest are burned annually, with yearly values ranging from 1375 to 22,951 hectares. Though less frequent than in savanna and grassland ecosystems, fires in forested areas have considerable ecological consequences, affecting biodiversity, carbon storage, and habitat structure, demonstrating the forests’ vulnerability to fire impacts within the Caatinga biome.

Pasture areas saw substantial fire activity, with a total of 1.9 million hectares burned over the 38 years, translating to an average of 49,403 hectares per year, with annual fluctuations between 11,163 and 175,900 hectares. Pasture fires are largely influenced by human activities, as fire is commonly used for land management. However, the extent of burning highlights the need for improved fire management practices to reduce widespread damage and manage the risks associated with anthropogenic fires.

Agricultural land was affected less frequently, with 78,960 hectares burned over the study period, averaging 2025 hectares annually. However, the use of fire for land management poses a risk of uncontrolled fires spreading, with a range of 184 to 7389 hectares burned in individual years. This suggests the need for sustainable fire management practices in agricultural landscapes to mitigate potential damage.

Similarly, the Mosaic of Uses, which refers to a mix of anthropogenic land uses, was significantly affected by fire activity, with 1.5 million hectares burned throughout the 38-year period. On average, 40,650 hectares are burned annually, with yearly variations from 9548 to 148,003 hectares. The fire impact in these mixed-use landscapes reflects diverse land management practices and highlights how human activities drive fire dynamics in the region.

When pasture, agriculture, and mosaics of uses are combined into the broader category of farming-related land uses, a total of 3.5 million hectares have burned over the study period, averaging 92,078 hectares per year. This substantial area indicates that farming activities, including pasture management and other agricultural practices, are major contributors to fire occurrence in the Caatinga, emphasizing the need for integrated fire management strategies across these land uses.

Wetlands, by contrast, experienced minimal fire activity, with only 34 hectares burned in total across 38 years, averaging less than 1 hectare annually. The high water content of wetlands likely makes them less susceptible to fires compared to other ecosystems, reinforcing their role as fire-resistant landscapes in the Caatinga biome.

Forest plantations recorded the lowest fire incidence, with only 233 hectares burned over the study period, demonstrating the effectiveness of fire prevention and management in these areas. The category labeled others, which includes various land use types, experienced 74,901 hectares burned over 38 years, with an annual average of 1921 hectares, indicating relatively low fire activity in these diverse landscapes.

3.5. Fire Frequency and Spatial Patterns

The analysis of overlapping maps of burned areas revealed regions within the Caatinga biome that experienced multiple fire events, indicating an increased fire frequency. The number of fire occurrences ranged from 1 to over 16, highlighting that specific locations were repeatedly affected by fire over the years (Figure 7).

As illustrated in Figure 7, a substantial proportion of the area (35%) experienced a single fire event between 1985 and 2023, indicating that many regions were only minimally affected. In contrast, 20% of the area experienced two fires, while 21% recorded three or four fires. Remarkably, only a small fraction (2%) of the area experienced fire events more than 16 times, highlighting the variability of fire occurrences across the landscape. This comprehensive analysis underscores the complexity of fire dynamics in the Caatinga, demonstrating that while many areas are subject to occasional fires, only select regions experience repeated fire events (Figure 7A,B).

The lower frequency of fires within the Caatinga biome is a complex phenomenon, influenced by a myriad of factors. These factors, including diverse land use patterns, varying vegetation composition, and local climate fluctuations, create a rich tapestry of physical and ecological conditions. This complexity results in a mosaic of fire-prone areas and more fire-resilient areas within the Caatinga biome, a fascinating aspect of our research that warrants further exploration.

Fire occurrences were more common in the western part of the Caatinga biome (Figure 7A). Still, factors beyond land use, vegetation, and the transition with the Cerrado biome may have also contributed to this pattern. In the eastern regions of the biome, human activities such as agriculture expansion, logging, and infrastructure development have also increased fire incidents. Understanding the spatial distribution of fire frequency in the Caatinga biome requires a comprehensive study of the complex interactions between natural and human-driven factors that influence fire occurrence.

3.6. Fire Dynamics in Subregions Within the Caatinga Biome

The analysis of accumulated mapped burn scars across 38 years within the Caatinga biome reveals an uneven distribution among the involved states. Bahia has the largest affected area, totaling 92.19 Mha, which accounts for approximately 36.88% of the cumulative total of 250.08 Mha. Following Bahia, Piauí presents 83.51 Mha, corresponding to 33.40% of the total. Ceará contributes 29.97 Mha (12.00%), while Pernambuco accumulates 13.63 Mha (5.44%) and Minas Gerais 12.67 Mha (5.07%). Paraíba records 10.80 Mha (4.31%), followed by Rio Grande do Norte with 5.16 Mha (2.06%). Alagoas and Sergipe exhibit the most minor affected areas, with 1.21 Mha (0.48%) and 0.94 Mha (0.38%), respectively. These findings highlight the predominance of burnings in Bahia and Piauí within the Caatinga biome, emphasizing the need for targeted regional strategies to mitigate the resulting environmental impacts.

The analysis of mapped burn scars across municipalities within the states defined by the Caatinga biome indicates a notable concentration of burned areas in specific locations. Among these, the municipality of Barra (BA) has the largest accumulated burned area, totaling 0.474 million hectares (Mha), representing approximately 4.3% of the total burned area recorded in the dataset. Following Barra, Pilão Arcado (BA) accounts for 0.310 Mha (2.8%), while Pimenteiras (PI) reports 0.222 Mha (2%), and Parnaguá (PI) contributes 0.217 Mha (1.9%). Muquém do São Francisco (BA) and Bom Jesus da Lapa (BA) register burned areas of 0.161 Mha (1.4%) and 0.126 Mha (1.1%), respectively. Additionally, the municipalities of São Miguel do Tapuio (PI), Canto do Buriti (PI), Curimatá (PI), and Morro Cabeça no Tempo (PI) each report burned areas of 0.1 Mha (1%). These top ten municipalities collectively represent around 21% of the total burned area documented in the dataset (see more details in Table S1 Supplementary Materials).

3.7. Accuracy Analysis

The evaluation of accuracy metrics for fire scar mapping in the Caatinga biome reveals consistent patterns across different years, shedding light on the strengths and weaknesses of the mapping process. The dataset encompasses overall accuracy, precision, recall, F1 score, and Jaccard index for 2007, 2013, 2014, 2015, and 2017. These metrics provide a comprehensive insight into the performance of fire scar detection. The Jaccard index, for instance, measures the similarity between two sets, in this case, the predicted and actual fire scar areas, offering valuable perspectives on the efficacy of mapping techniques over time.

The overall accuracy in 2007 reached 0.94, signifying a high rate of correct classifications. The precision was notably high at 0.97, indicating that almost all identified fire scars were true positives (Table 2). However, the recall was comparatively lower at 0.87, suggesting some actual fire scars were not identified. The F1 score, which balances precision and recall, was 0.91, while the Jaccard index stood at 0.84 (Table 2). This performance reflects a robust overall outcome with a slight compromise between identifying all fire scars and ensuring the correctness of detected scars, highlighting the trade-offs involved in the mapping process (Table 2).

In 2013, we observed the highest performance across all metrics. The overall accuracy was 0.98, the highest among the observed years, indicating nearly perfect classification. Precision and recall stood at 0.98 and 0.95, respectively, demonstrating a balanced and impressively accurate identification of fire scars (Table 2). The F1 score peaked at 0.96, and the Jaccard index was 0.93 (Table 2). This outstanding performance suggests that the mapping techniques used in 2013 played a crucial role in effectively detecting and accurately classifying fire scars in the Caatinga biome (Table 2).

The data from 2014 showed some variability in the metrics. While the overall accuracy remained high at 0.97, the precision decreased significantly to 0.82 (Table 2). This decline indicates an increase in false positives, where non-fire areas were mistakenly identified as fire scars (Table 2). The recall rate was 0.85, consistent with previous years, and the F1 score stood at 0.83, highlighting the imbalance between precision and recall. The Jaccard index notably dropped to 0.74, suggesting reduced overlap between predicted and actual fire scar areas. This year serves as a stark reminder of the challenges in achieving high precision while maintaining high recall in fire scar mapping, a struggle that many in the field can empathize with and that underscores the complexity of the task (Table 2).

In 2015, the overall accuracy decreased to 0.92, representing the lowest value observed among the years under study. The precision remained relatively high at 0.94, while the recall was 0.85 (Table 2), consistent with previous years. The F1 score was 0.88, and the Jaccard index was 0.80 (Table 2). This year highlighted a persistent issue with missing some fire scars (resulting in lower recall) but effectively identifying those detected (resulting in high precision) (Table 2).

In 2017, mapping performance significantly improved, achieving an overall accuracy of 0.96. Precision was notably high at 0.97, and recall showed improvement at 0.93, indicating a well-balanced and effective detection of fire scars. The F1 score reached 0.95, and the Jaccard index stood at 0.91 (Table 2), demonstrating a high overlap between the predicted and actual fire scar areas. These enhancements reflect the progress in mapping techniques and data accuracy in recent years, inspiring confidence in the future of fire scar mapping and the potential for even more accurate and effective techniques.

3.8. Correlation Between Fire Burn Scars and Climate Variables

The spatial analysis of correlations between accumulated burn scars and three key climatic variables—precipitation accumulation, climate water deficit, and reference evapotranspiration—reveals distinct patterns across northeastern Brazil, suggesting complex interactions between climate conditions and fire incidence.

The analysis of the correlation between accumulated burn scars and the 30-year average precipitation accumulation (Figure 8A) reveals a predominance of negative correlations, particularly in the southern portion of the northeastern region, encompassing areas like western Bahia and northern Minas Gerais. This spatial pattern suggests that regions with lower average precipitation are more susceptible to fire occurrences. These arid or semi-arid areas exhibit a robust inverse relationship between precipitation levels and fire incidence. In contrast, positive correlations, where higher precipitation coincides with increased burn scars, are infrequent and primarily observed in isolated coastal areas, such as the Recôncavo region in Bahia. Overall, the spatial configuration in Figure 8A is dominated by negative correlations or a lack of significant correlations in central areas, reinforcing the notion that low precipitation may predispose certain regions to a higher frequency of fire events.

Figure 8B, which examines the correlation between burn scars and the 30-year average climate water deficit, reveals a notable preponderance of positive correlations across large swaths of the interior northeast, particularly in areas such as western Pernambuco, the semi-arid zone of Paraíba, and parts of Piauí and Ceará. These regions, known to experience high climate water deficits, exhibit a robust association between increased water stress and the prevalence of fire events. This spatial pattern suggests that areas facing greater water scarcity are more susceptible to fires, potentially due to reduced vegetation moisture and heightened vulnerability to ignition. In contrast, negative correlations are infrequent and scattered, implying that water deficit and fire occurrence are predominantly positively correlated in most areas examined.

The analysis of the correlation between burn scars and the 30-year average reference evapotranspiration, as depicted in Figure 8C, reveals a substantial prevalence of positive associations in several regions. Notably, these positive correlations are most concentrated in centralwestern Bahia, southern Pernambuco, and northern Minas Gerais. In these areas characterized by high evapotranspiration rates, there is a marked relationship between increased atmospheric water demand and fire events. This pattern suggests that the more significant moisture loss from soil and vegetation in these regions may facilitate fire spread. Conversely, negative correlations are scarce and appear sporadically, primarily in the more humid coastal areas. The widespread positive correlations found in this spatial configuration suggest that regions with high evapotranspiration, dryness, and low vegetation moisture appear to favor fire incidence.

The spatial patterns reveal a very sparse positive correlation between fire incidence and climatic factors that intensify drought conditions, such as water deficit and high evapotranspiration, especially in semi-arid interior regions. In contrast, the correlation with precipitation accumulation is predominantly negative, indicating that areas with lower rainfall mainly in the interior of the northeast and northern Minas Gerais are more susceptible to fires. These findings emphasize the critical role of climate-induced water scarcity in fire dynamics, particularly in regions with a high climate water deficit and elevated evapotranspiration, which seem to predispose these areas to an increased risk of fires. However, in our study, it must be evident that they are predominant determinants from a spatial point of view for the presence of fire or its dynamics, as they do not present a particular dominance in terms of significance as described above.

4. Discussion

Studying fire scars across different land use categories from 1985 to 2023 provides valuable insights into spatial distribution and long-term impact burn areas in the Caatinga biome. The dataset includes various land use classes and annual records of fire-affected areas. These findings are crucial for comprehending fire dynamics and their effects on land use practices and ecosystem health within the semi-arid Caatinga biome in northeastern Brazil.

The relationship between fire and the evolutionary history of Caatinga vegetation suggests that fire did not play a significant role in shaping its flora. According to [18], the Caatinga biome is classified as fire-independent, meaning that fire was not a key factor in the evolution or adaptation of its plant species. Natural fire events in the Caatinga are expected to be relatively rare [31], largely due to the discontinuous nature of the fuel, which limits the spread of fire [59], and the low incidence of lightning events [60], which are typically a primary cause of natural fires in other ecosystems.

However, the present study’s examination of the temporal patterns of burned areas from 1985 to 2023 reveals a complex interaction between climatic, environmental, and human-influenced factors that shape wildfire dynamics in the Caatinga. The data align with known fire patterns, where prolonged dry seasons and human activities significantly contribute to fire occurrences [16,34]. Research by [25] has demonstrated that fire occurrences are highly influenced by temperature and precipitation during the dry season, with higher temperatures and reduced rainfall exacerbating fire risk.

In terms of causality, most fires in the region are anthropogenic in origin, often associated with farming practices and land management strategies. Consequently, the current fire activity in the Caatinga represents a disruption of its natural ecological processes rather than a continuation of its evolutionary history. This presents a significant threat to the biome’s species, many of which are not adapted to frequent fire disturbances, further amplifying the risk of ecological degradation and biodiversity loss. Simulations have estimated that natural regeneration after anthropogenic fires requires at least 50 years [31], sometimes leading to desertification [61,62].

The present research analysis reveals substantial year-to-year variability in fire disturbances, with some years experiencing significantly larger burned areas. Notably, 2021 recorded the highest total burned area at 899,109 hectares. This pronounced peak aligns with previous studies linking extreme fire events to climatic anomalies such as severe droughts and prolonged heatwaves, which increase fuel loads and fire risks [16]. These conditions, often associated with climate change, have been shown to heighten wildfire occurrences in dryland ecosystems, including the Caatinga [30]. In contrast, 1985 saw the smallest burned area, totaling 183,602 hectares. The lower fire activity during this period may reflect more favorable climatic conditions, such as adequate rainfall and moderate temperatures, alongside reduced human-induced ignition sources.

The predominance of small and medium fire scars across most years suggests that fire events in the Caatinga are often limited in scale. This could be attributed to the biome‘s fragmented nature, which restricts fire spread, or to frequent small-scale fires that prevent the accumulation of large fuel loads needed for larger fires. The presence of moderate-sized fire scars indicates that while many fires remain small, a significant number still grow to intermediate sizes. However, from the early 2000s onward, there has been a marked increase in burned areas, indicating a shift in the fire regime. Larger and more frequent fire events have become the norm, particularly in years such as 2010, and 2015.

The extensive analysis of accumulated burn scars in the Caatinga biome over the past 38 years reveals a significantly uneven distribution of affected areas among the participating states. The pronounced concentration of burn scars, primarily in Bahia and Piauí, highlights critical regional disparities within the biome. These findings may reflect the states differing socio-economic conditions and land-use practices. For instance, the extensive agricultural activities and land management policies in Bahia may contribute to the higher incidence of burn scars. Likewise, Piauí reliance on fire-based land-clearing techniques may further exacerbate the extent of burned areas. This regional imbalance underscores the need for tailored mitigation strategies that address each state’s specific drivers of burning activities. Collaborative efforts involving local governments, communities, and environmental organizations are essential to develop sustainable land management practices that minimize burn scars and enhance ecosystem resilience. Prioritizing practical mitigation efforts in these regions is crucial to curtailing further degradation and promoting the restoration of the fragile ecosystems within the Caatinga.

The municipal-level analysis reveals a concentrated pattern of burn scars, with the top ten municipalities accounting for approximately 21% of the total burned area. Within this context, the concentration of burn scars in critical regions such as Barra (BA) and Pilão Arcado (BA) underscores the influence of localized factors on fire incidence. These municipalities could serve as focal points for implementing targeted interventions, including enhancements to fire management infrastructure, promoting alternative agricultural practices, and increasing community education on fire prevention. Furthermore, the significant share of burn scars in these top ten municipalities highlights potential hotspots that necessitate immediate attention to avert further environmental degradation. Future research should delve into these areas’ socio-economic and environmental conditions to inform more effective, context-specific policy measures. Addressing these localized challenges is vital for preserving the biodiversity of the Caatinga biome and sustaining its ecological functions. Essential steps toward reducing the prevalence of burn scars and mitigating their long-term environmental impacts within the Caatinga biome include enhancing community engagement, improving fire management practices, and implementing sustainable land use policies.

In the present study, the largest fire scar areas, in terms of total hectares affected, were concentrated between September and December, aligning with the burning season identified by [16] for the Caatinga biome. Interestingly, this pattern is consistent with the findings of [63], who also observed that October marked the peak of fire activity, despite not being among the driest months. The authors [63] suggested that the biomass produced during the rainy season requires several months after the rains to dry out sufficiently and become fire-prone fuel, resulting in large fire events later in the year. The authors of [64] reported the highest probability of burned areas occurring from August to September in the Amazon, Atlantic Forest, and Cerrado biomes; this reflects the diversity of fire regimes across ecosystems. The high fire activity observed in October for the Caatinga underscores the importance of tailored mitigation efforts, including controlled burns and community awareness programs, aimed at preventing accidental ignitions and reducing fire risks.

Our results demonstrate that fire events are notably concentrated in savanna vegetation, particularly at its interfaces with the Cerrado. This pattern aligns with previous studies, such as that by [16], who documented high fire incidence in the ecotonal areas between savanna and forested regions in the Cerrado biome. Similarly, [25] found that savanna formations and rocky outcrops are major drivers of fire in the Caatinga. The concentration of fire in these areas underscores the importance of targeted fire management strategies in ecotonal zones, which are particularly vulnerable to frequent burning.

The frequent fires in Caatinga’s savanna and pasture areas degrade native vegetation, reduce biodiversity, and alter soil properties. These fires often stem from land-clearing practices for agricultural expansion or pasture maintenance, underscoring the need for sustainable land management approaches that balance productivity and ecological conservation. Implementing strategies that promote integrated land-use planning, fire-smart agricultural techniques, and the restoration of degraded areas can help mitigate the adverse impacts of fire on Caatinga’s natural ecosystems. While forest and wetland areas within the Caatinga biome experience relatively lower fire incidence, they can still be significantly impacted by the broader fire dynamics in the region. Caatinga’s forested ecosystems provide critical wildlife habitats and serve as important carbon sinks [54], making their protection from fire a paramount concern for maintaining the overall ecological integrity and resilience of the biome.

The variability in fire activity, both year-to-year and seasonally, along with its spatial concentration in specific areas, underscores the need for ongoing monitoring and adaptive fire management strategies in the Caatinga biome. This dynamic fire regime demands the engagement of scientists, policymakers, and local communities to address the escalating fire risks effectively. Tailored management approaches should integrate traditional knowledge, modern fire management practices, and sustainable land-use policies to mitigate the impacts of fire and enhance the resilience of this unique ecosystem.

Finally, it is important to emphasize that Caatinga’s high susceptibility to fires is a critical factor in its ecological dynamics, underscoring the need for precise fire scar mapping. This mapping is not only a scientific endeavor but also a crucial tool for understanding fire patterns, assessing ecological impacts, and implementing effective conservation measures. Therefore, accurate fire scar mapping is a critical scientific effort, serving as a vital resource for ecological assessment and conservation initiatives. In this context, the accuracy metrics analyzed over several years reveal ongoing efforts to refine mapping techniques while highlighting the challenges of balancing precision and recall.

The strong performance observed in this study, particularly in the years 2013 and 2017, can be attributed to advancements in remote sensing technologies and algorithms that significantly improved fire scar detection [16]. Conversely, the lower accuracy recorded in 2014, marked by a higher occurrence of false positives, underscores the necessity for improved methodologies to minimize such inaccuracies. Addressing these trends is crucial for developing resilient fire monitoring systems that support effective management and conservation initiatives in the Caatinga biome.

Examining the accuracy metrics of fire scar mapping over multiple years provides essential insights into the efficacy of various mapping techniques. The consistently high overall accuracy suggests the reliability of current methods, while fluctuations in precision and recall indicate specific areas that require improvement. These findings are fundamental for advancing fire monitoring and management practices in the Caatinga, ultimately contributing to more effective conservation efforts and a deeper understanding of this unique ecosystem.

The comprehensive analysis of fire scars within the Caatinga biome from 1985 to 2023 offers critical practical applications for enhancing land management and conservation efforts. By identifying hotspots with high concentrations of burn scars, particularly in Bahia and Piauí, resources can be prioritized for targeted fire prevention and management initiatives such as controlled burns, the establishment of firebreaks, and the deployment of rapid response teams. Integrating advanced remote sensing technologies with ground-based monitoring systems will improve the accuracy and timeliness of fire detection, enabling more efficient allocation of resources and timely interventions.

Additionally, developing sustainable land-use policies that promote alternative agricultural practices like agroforestry and no-till farming can significantly reduce anthropogenic fire activities. Engaging local communities through education and participatory management programs is essential for fostering fire-smart behaviors and enhancing resilience against wildfires. Furthermore, incorporating traditional ecological knowledge with modern scientific approaches will create comprehensive fire management frameworks. Collaborative efforts among policymakers, scientists, and local stakeholders are crucial to implementing these strategies effectively, thereby preserving the biodiversity and ecological integrity of the Caatinga biome and ensuring its sustainability amidst increasing fire disturbances.

The findings underscore the intricate and spatially sparse relationship between fire incidence and climatic variables in the semi-arid and arid regions of northeastern Brazil, where climate-driven water scarcity plays a role in influencing fire dynamics. The observed patterns are notably dispersed, lacking a dominant or consistent trend across any specific region, which underscores the complex and heterogeneous nature of fire-climate interactions in this area. Despite this overall sparseness, certain localized areas exhibit significant correlations that hint at possible interactions between climate conditions and fire occurrences.

The precipitation patterns in the Caatinga biome exhibit a clear seasonality, with distinct wet and dry periods that influence vegetation growth and, consequently, fire dynamics. This seasonal rainfall variation creates cycles of biomass accumulation, which, once dried, serves as fuel for fires during dry months. Additionally, monthly precipitation anomalies reveal fluctuations from the historical average, indicating periods of unusually high or low rainfall that either mitigate or increase fire risks. Positive anomalies in rainfall can reduce fire likelihood by maintaining vegetation moisture, while negative anomalies heighten susceptibility by drying out vegetation (see Supplementary Materials Figure S1).

Together, the seasonality and anomalies in precipitation demonstrate the complex interaction between climate and fire dynamics in the Caatinga. These findings underscore the importance of understanding localized climate factors, as both typical and atypical rainfall patterns significantly impact fire potential. This knowledge can guide better fire management and risk assessment strategies within the biome (see Supplementary Materials Figure S1).

The predominance of negative correlations between precipitation accumulation and burn scars (Figure 8A) in some regions suggests that areas with lower rainfall are particularly vulnerable to fires, supporting theories that reduced moisture availability limits vegetation’s ability to withstand ignition and recover after burning. This relationship is further intensified by the presence of positive correlations between burn scars and climate water deficit (Figure 8B) and reference evapotranspiration (Figure 8C) in select areas within the interior northeast. These correlations, while spatially limited, highlight how increased water stress, whether due to low precipitation or high evapotranspiration, creates conditions conducive to fire spread by reducing vegetation moisture and increasing susceptibility to ignition.

Such sparse yet significant patterns align with global studies indicating that prolonged drought and high evapotranspiration often correlate with higher fire activity, particularly in ecosystems lacking resilience to extreme climate conditions. However, the absence of a uniform or dominant correlation pattern across the study area suggests that additional local factors, such as land use, vegetation type, and human activity, may play crucial roles in determining fire occurrence. While water scarcity driven by climate variables emerges as an influential factor, these variables do not solely account for fire dynamics in northeastern Brazil. Thus, climate-induced water stress should be considered within a broader framework that incorporates ecological, socio-economic, and anthropogenic influences to effectively assess and mitigate fire risks in this complex and variable landscape.

5. Conclusions

Analyzing fire scars across various land use categories in the Caatinga biome from 1985 to 2023 offers valuable insights into the complex dynamics of wildfires in this unique semi-arid ecosystem. The research reveals significant annual fluctuations in burned areas, with notable fire occurrences in specific years influenced by weather conditions such as severe droughts and prolonged heatwaves. The growing prevalence of burned areas highlights the increasing pressure on the Caatinga biome, further exacerbated by climate change and shifts in land use practices.

Continuous monitoring and flexible management strategies are essential for effectively protecting the Caatinga ecosystem from the harmful effects of wildfires. This requires active participation from all stakeholders, including researchers, environmentalists, policymakers, and local communities. Tailored approaches, such as comprehensive land-use planning, fire-resilient agricultural practices, and the restoration of degraded areas, can help reduce fire-related damage and improve the ecological resilience of the Caatinga biome.

The findings of this study clarify the intricate interplay of climatic, environmental, and human factors that influence wildfire dynamics, thus guiding the development of effective fire control measures to safeguard this unique ecosystem. The findings emphasize the need for ongoing monitoring and adaptable approaches to address the increasing wildfire risks in the Caatinga biome. Collaboration among various stakeholders, including scientists, policymakers, and local communities, is crucial for developing tailored solutions that promote sustainable land use, fire-resilient agricultural practices, and the restoration of affected areas. Addressing the root causes of fire disturbances and enhancing the ecological resilience of the Caatinga can help preserve this unique and invaluable ecosystem for present and future generations.

The notable annual variations in burned areas, characterized by intense fire occurrences in specific years, underscore the vulnerability of the Caatinga to fire impacts and the pressing need for adequate fire management techniques. The limited occurrence of extensive fire scars in the Caatinga biome can be attributed to a combination of ecological and environmental factors. The semi-arid climate of the biome, characterized by distinct wet and dry seasons, influences the dynamics and spread of wildfires. Additionally, human activities, such as unsustainable land use changes and inadequate fire control practices, can exacerbate the frequency and intensity of fires in the Caatinga. Widespread fire scars are often associated with severe climatic conditions or prolonged droughts, which facilitate the proliferation of flames across larger areas.

Crafting adaptive, targeted fire management strategies that harmonize traditional and contemporary methods is essential for bolstering the resilience of the Caatinga biome. Engaging all relevant stakeholders is pivotal to mitigating fire risks, safeguarding the biome‘s unique biodiversity, and supporting local communities.

Supplementary Materials

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/fire7120437/s1, Supplementary Materials Table S1. The table describes the mapped Accumulated Burned Area for each municipality within the Caatinga biome over the time series from 1985 to 2023. Supplementary Materials Figure S1. The figure depicts the temporal pattern of precipitation and monthly anomaly values for the entire Caatinga biome, based on data from the TerraClimate.

Author Contributions

Conceptualization, W.J.S.F.R., R.N.V., M.M.M.d.S., S.G.D., D.P.C., S.M, C.L.C., L.d.S.B., M.O., J.F.-F., W.V.d.S., V.L.S.A., A.A.C.A., R.O.F.R. and N.A.S.; methodology, W.J.S.F.R., R.N.V., M.M.M.d.S., S.G.D., D.P.C., S.M, C.L.C., L.d.S.B., M.O., J.F.-F., W.V.d.S., V.L.S.A., A.A.C.A., R.O.F.R. and N.A.S.; software execution, W.J.S.F.R., R.N.V., D.P.C., S.M, C.L.C., L.d.S.B., M.O., J.F.-F., W.V.d.S., V.L.S.A., A.A.C.A., R.O.F.R., N.A.S. and S.G.D. writing—original draft preparation, W.J.S.F.R., R.N.V., M.M.M.d.S., S.G.D., D.P.C., S.M, C.L.C., L.d.S.B., M.O., J.F.-F., W.V.d.S., V.L.S.A., A.A.C.A., R.O.F.R. and N.A.S.; writing—review and editing, W.J.S.F.R., R.N.V. and M.M.M.d.S. supervision, W.J.S.F.R. and R.N.V. funding acquisition, W.J.S.F.R., L.d.S.B., M.O. and J.F.-F. All authors have read and agreed to the published version of the manuscript.

Funding

W.J.S.F.R. was supported by a CNPQ research fellowship under Process #314954/2021-0 and the Prospecta 4.0–CNPQ research grant under Process #407907/2022-0. D.P.C. was financially supported by the Bahia State Research Foundation (FAPESB) under grant #BOL 0457/2019 and by CAPES/CAPES/PRINT through Edital n° 41/2017. D.P.C, R.N.V., and W.J.S.F.R. were supported by the INCT IN-TREE for Technology in Interdisciplinary and Transdisciplinary Studies in Ecology and Evolution CNPQ research grant under n° 465767/2014-1. D.P.C, R.N.V., S.G.D. and W.J.S.F.R. were supported by the WRI subgrant to WRI Brasil n° 73054 related to the Land and Carbon Lab platform.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

The original contributions presented in the study are included in the article/Supplementary Material, further inquiries can be directed to the corresponding author.

Acknowledgments

We appreciate comments and suggestions from the anonymous reviewers that helped improve the quality and presentation of the manuscript.

Conflicts of Interest

The authors declare no conflicts of interest.

References

Flannigan, M.; Cantin, A.S.; De Groot, W.J.; Wotton, M.; Newbery, A.; Gowman, L.M. Global Wildland Fire Season Severity in the 21st Century. For. Ecol. Manag. 2013, 294, 54–61. [Google Scholar] [CrossRef]
Sayedi, S.S.; Abbott, B.W.; Vannière, B.; Leys, B.; Colombaroli, D.; Romera, G.G.; Słowiński, M.; Aleman, J.C.; Blarquez, O.; Feurdean, A.; et al. Assessing Changes in Global Fire Regimes. Fire Ecol. 2024, 20, 18. [Google Scholar] [CrossRef]
Cunningham, C.X.; Williamson, G.J.; Bowman, D.M.J.S. Increasing Frequency and Intensity of the Most Extreme Wildfires on Earth. Nat. Ecol. Evol. 2024, 8, 1420–1425. [Google Scholar] [CrossRef]
Jones, B.A. Are We Underestimating the Economic Costs of Wildfire Smoke? An Investigation Using the Life Satisfaction Approach. J. For. Econ. 2017, 27, 80–90. [Google Scholar] [CrossRef]
Johnston, F.H.; Borchers-Arriagada, N.; Morgan, G.G.; Jalaludin, B.; Palmer, A.J.; Williamson, G.J.; Bowman, D.M.J.S. Unprecedented Health Costs of Smoke-Related PM2.5 from the 2019–2020 Australian Megafires. Nat. Sustain. 2021, 4, 42–47. [Google Scholar] [CrossRef]
Butt, E.W.; Conibear, L.; Reddington, C.L.; Darbyshire, E.; Morgan, W.T.; Coe, H.; Artaxo, P.; Brito, J.; Knote, C.; Spracklen, D.V. Large Air Quality and Human Health Impacts Due to Amazon Forest and Vegetation Fires. Environ. Res. Commun. 2020, 2, 095001. [Google Scholar] [CrossRef]
Kelley, D.I.; Bistinas, I.; Whitley, R.; Burton, C.; Marthews, T.R.; Dong, N. How Contemporary Bioclimatic and Human Controls Change Global Fire Regimes. Nat. Clim. Chang. 2019, 9, 690–696. [Google Scholar] [CrossRef]
Bowman, D.M.J.S.; Balch, J.K.; Artaxo, P.; Bond, W.J.; Carlson, J.M.; Cochrane, M.A.; D’Antonio, C.M.; DeFries, R.S.; Doyle, J.C.; Harrison, S.P.; et al. Fire in the Earth System. Science 1979 2009, 324, 481–484. [Google Scholar] [CrossRef] [PubMed]
Morton, D.C.; Defries, R.S.; Randerson, J.T.; Giglio, L.; Schroeder, W.; Van Der Werf, G.R. Morton Agricultural Intensification Increases Deforestation Fire Activity in Amazonia. Glob. Chang. Biol. 2008, 14, 2262–2275. [Google Scholar] [CrossRef]
Balch, J.K.; Bradley, B.A.; Abatzoglou, J.T.; Chelsea Nagy, R.; Fusco, E.J.; Mahood, A.L. Human-Started Wildfires Expand the Fire Niche across the United States. Proc. Natl. Acad. Sci. USA 2017, 114, 2946–2951. [Google Scholar] [CrossRef]
Nepstad, D.; Carvalho, G.; Barros, A.C.; Alencar, A.; Capobianco, J.P.; Bishop, J.; Moutinho, P.; Lefebvre, P.; Silva, U.L.; Prins, E. Road Paving, Fire Regime Feedbacks, and the Future of Amazon Forests. For. Ecol. Manag. 2001, 154, 395–407. [Google Scholar] [CrossRef]
De Oliveira-Júnior, J.F.; Mendes, D.; Correia Filho, W.L.F.; da Silva Junior, C.A.; de Gois, G.; Jardim, A.M.d.R.F.; da Silva, M.V.; Lyra, G.B.; Teodoro, P.E.; Pimentel, L.C.G.; et al. Fire Foci in South America: Impact and Causes, Fire Hazard and Future Scenarios. J. S. Am. Earth Sci. 2021, 112, 103623. [Google Scholar] [CrossRef]
Li, P.; Xiao, C.; Feng, Z.; Li, W.; Zhang, X. Occurrence Frequencies and Regional Variations in Visible Infrared Imaging Radiometer Suite (VIIRS) Global Active Fires. Glob. Chang. Biol. 2020, 26, 2970–2987. [Google Scholar] [CrossRef]
Lapola, D.M.; Pinho, P.; Barlow, J.; Aragão, L.E.O.C.; Berenguer, E.; Carmenta, R.; Liddy, H.M.; Seixas, H.; Silva, C.V.J.; Silva, C.H.L.; et al. The Drivers and Impacts of Amazon Forest Degradation. Science 1979 2023, 379, eabp8622. [Google Scholar] [CrossRef]
Cochrane, M.A.; Cochrane, M.A.; Ryan, K.C. Fire and Fire Ecology: Concepts and Principles. In Tropical Fire Ecology; Springer: Berlin/Heidelberg, Germany, 2009; pp. 25–62. [Google Scholar]
Alencar, A.A.C.; Arruda, V.L.S.; da Silva, W.V.; Conciani, D.E.; Costa, D.P.; Crusco, N.; Duverger, S.G.; Ferreira, N.C.; Franca-Rocha, W.; Hasenack, H.; et al. Long-Term Landsat-Based Monthly Burned Area Dataset for the Brazilian Biomes Using Deep Learning. Remote Sens. 2022, 14, 2510. [Google Scholar] [CrossRef]
Simon, M.F.; Pennington, T. Evidence for Adaptation to Fire Regimes in the Tropical Savannas of the Brazilian Cerrado. Int. J. Plant Sci. 2012, 173, 711–723. [Google Scholar] [CrossRef]
Hardesty, J.; Myers, R.; Fulks, W. Fire, Ecosystems and People: A Preliminary Assessment of Fire as a Global Conservation Issue. Fire Manag. 2005, 22, 78–87. [Google Scholar]
Pivello, V.R.; Vieira, I.; Christianini, A.V.; Ribeiro, D.B.; da Silva Menezes, L.; Berlinck, C.N.; Melo, F.P.L.; Marengo, J.A.; Tornquist, C.G.; Tomas, W.M.; et al. Understanding Brazil’s Catastrophic Fires: Causes, Consequences and Policy Needed to Prevent Future Tragedies. Perspect. Ecol. Conserv. 2021, 19, 233–255. [Google Scholar] [CrossRef]
Gatti, L.V.; Basso, L.S.; Miller, J.B.; Gloor, M.; Gatti Domingues, L.; Cassol, H.L.G.; Tejada, G.; Aragão, L.E.O.C.; Nobre, C.; Peters, W.; et al. Amazonia as a Carbon Source Linked to Deforestation and Climate Change. Nature 2021, 595, 388–393. [Google Scholar] [CrossRef]
de Santana, M.M.M.; de Vasconcelos, R.N.; Mariano Neto, E.; da Franca Rocha, W.d.J.S. Machine Learning Model Reveals Land Use and Climate’s Role in Amazon Wildfires: Present and Future Scenarios. Fire 2024, 7, 338. [Google Scholar] [CrossRef]
Le Page, Y.; Morton, D.; Hartin, C.; Bond-Lamberty, B.; Pereira, J.M.C.; Hurtt, G.; Asrar, G. Synergy between Land Use and Climate Change Increases Future Fire Risk in Amazon Forests. Earth Syst. Dyn. 2017, 8, 1237–1246. [Google Scholar] [CrossRef]
Siyum, Z.G. Tropical Dry Forest Dynamics in the Context of Climate Change: Syntheses of Drivers, Gaps, and Management Perspectives. Ecol. Process. 2020, 9, 25. [Google Scholar] [CrossRef]
Oliveira, U.; Soares-Filho, B.; Bustamante, M.; Gomes, L.; Ometto, J.P.; Rajão, R. Determinants of Fire Impact in the Brazilian Biomes. Front. For. Glob. Chang. 2022, 5, 735017. [Google Scholar] [CrossRef]
Martins, S.F.S.; dos Santos, A.M.; da Silva, C.F.A.; Rudke, A.P.; Alvarado, S.T.; Melo, J.L.d.S. The Drivers of Fire in the Caatinga Biome in Brazil. For. Ecol. Manag. 2024, 572, 122260. [Google Scholar] [CrossRef]
da Silva Junior, J.A.; Pacheco, A.d.P. Fire Analysis in the Caatinga Environment from Landsat-8 Images, Enhanced Vegetation Index and Analysis by the Main Components. Cienc. Florest. 2021, 31, 417–439. [Google Scholar] [CrossRef]
de Araújo, F.M.; Ferreira, L.G.; Arantes, A.E. Distribution Patterns of Burned Areas in the Brazilian Biomes: An Analysis Based on Satellite Data for the 2002–2010 Period. Remote Sens. 2012, 4, 1929. [Google Scholar] [CrossRef]
Lucas, F.M.F.; Araujo, E.C.G.; Fiedler, N.C.; Santana, J.A.d.S.; Tetto, A.F. Scientific Gaps on Forest Fires in Brazilian Protected Areas. For. Ecol. Manag. 2023, 529, 120739. [Google Scholar] [CrossRef]
Bezerra, J.S.; Arroyo-Rodríguez, V.; Tavares, J.M.; Leal, A.; Leal, I.R.; Tabarelli, M. Drastic Impoverishment of the Soil Seed Bank in a Tropical Dry Forest Exposed to Slash-and-Burn Agriculture. For. Ecol. Manag. 2022, 513, 120185. [Google Scholar] [CrossRef]
Viegas, L.M.D.; Sales, L.; Hipólito, J.; Amorim, C.; de Pereira, E.J.; Ferreira, P.; Folta, C.; Ferrante, L.; Fearnside, P.; Malhado, A.C.M.; et al. We’re Building It up to Burn It down: Fire Occurrence and Fire-Related Climatic Patterns in Brazilian Biomes. PeerJ 2022, 10, e14276. [Google Scholar] [CrossRef]
Althoff, T.D.; Menezes, R.S.C.; de Carvalho, A.L.; de Siqueira Pinto, A.; Santiago, G.A.C.F.; Ometto, J.P.H.B.; von Randow, C.; de Sá Barretto Sampaio, E.V. Climate Change Impacts on the Sustainability of the Firewood Harvest and Vegetation and Soil Carbon Stocks in a Tropical Dry Forest in Santa Teresinha Municipality, Northeast Brazil. For. Ecol. Manag. 2016, 360, 367–375. [Google Scholar] [CrossRef]
de Moraes, C.A.; de Oliveira, M.A.T.; Behling, H. Late Holocene Climate Dynamics and Human Impact Inferred from Vegetation and Fire History of the Caatinga, in Northeast Brazil. Rev. Palaeobot. Palynol. 2020, 282, 104299. [Google Scholar] [CrossRef]
Antongiovanni, M.; Venticinque, E.M.; Matsumoto, M.; Fonseca, C.R. Chronic Anthropogenic Disturbance on Caatinga Dry Forest Fragments. J. Appl. Ecol. 2020, 57, 2064–2074. [Google Scholar] [CrossRef]
de Oliveira-Júnior, J.F.; Shah, M.; Abbas, A.; Correia Filho, W.L.F.; da Silva Junior, C.A.; Santiago, D.D.B.; Teodoro, P.E.; Mendes, D.; de Souza, A.; Aviv-Sharon, E.; et al. Spatiotemporal Analysis of Fire Foci and Environmental Degradation in the Biomes of Northeastern Brazil. Sustainability 2022, 14, 6935. [Google Scholar] [CrossRef]
de Santana, M.M.M.; Mariano-Neto, E.; de Vasconcelos, R.N.; Dodonov, P.; Medeiros, J.M.M. Mapping the Research History, Collaborations and Trends of Remote Sensing in Fire Ecology. Scientometrics 2021, 126, 1359–1388. [Google Scholar] [CrossRef]
Collins, L.; Griffioen, P.; Newell, G.; Mellor, A. The Utility of Random Forests for Wildfire Severity Mapping. Remote Sens. Environ. 2018, 216, 374–384. [Google Scholar] [CrossRef]
Abid, F. A Survey of Machine Learning Algorithms Based Forest Fires Prediction and Detection Systems. Fire Technol. 2021, 5, 559–590. [Google Scholar] [CrossRef]
Mohajane, M.; Costache, R.; Karimi, F.; Bao Pham, Q.; Essahlaoui, A.; Nguyen, H.; Laneve, G.; Oudija, F. Application of Remote Sensing and Machine Learning Algorithms for Forest Fire Mapping in a Mediterranean Area. Ecol. Indic. 2021, 129, 107869. [Google Scholar] [CrossRef]
Shao, Y.; Feng, Z.; Sun, L.; Yang, X.; Li, Y.; Xu, B.; Chen, Y. Mapping China’s Forest Fire Risks with Machine Learning. Forests 2022, 13, 856. [Google Scholar] [CrossRef]
LeCun, Y.; Bengio, Y.; Hinton, G. Deep Learning. Nature 2015, 521, 436–444. [Google Scholar] [CrossRef]
Learning, D. Deep Learning—Goodfellow. Nature 2016, 26. [Google Scholar]
Langford, Z.; Kumar, J.; Hoffman, F. Wildfire Mapping in Interior Alaska Using Deep Neural Networks on Imbalanced Datasets. In Proceedings of the IEEE International Conference on Data Mining Workshops, ICDMW, Singapore, 17–20 November 2018. [Google Scholar] [CrossRef]
Saeed, F.; Paul, A.; Karthigaikumar, P.; Nayyar, A. Convolutional Neural Network Based Early Fire Detection. Multimed. Tools Appl. 2020, 79, 9083–9099. [Google Scholar] [CrossRef]
Arruda, V.L.S.; Piontekowski, V.J.; Alencar, A.; Pereira, R.S.; Matricardi, E.A.T. An Alternative Approach for Mapping Burn Scars Using Landsat Imagery, Google Earth Engine, and Deep Learning in the Brazilian Savanna. Remote Sens. Appl. 2021, 22, 100472. [Google Scholar] [CrossRef]
Gholamnia, K.; Nachappa, T.G.; Ghorbanzadeh, O.; Blaschke, T. Comparisons of Diverse Machine Learning Approaches for Wildfire Susceptibility Mapping. Symmetry 2020, 12, 604. [Google Scholar] [CrossRef]
Gorelick, N.; Hancher, M.; Dixon, M.; Ilyushchenko, S.; Thau, D.; Moore, R. Remote Sensing of Environment Google Earth Engine: Planetary-Scale Geospatial Analysis for Everyone. Remote Sens. Environ. 2017, 202, 18–27. [Google Scholar] [CrossRef]
Olson, D.M.; Dinerstein, E.; Wikramanayake, E.D.; Burgess, N.D.; Powell, G.V.N.; Underwood, E.C.; D’Amico, J.A.; Itoua, I.; Strand, H.E.; Morrison, J.C.; et al. Terrestrial Ecoregions of the World: A New Map of Life on Earth. Bioscience 2001, 51, 933–938. [Google Scholar] [CrossRef]
De Albuquerque, U.P.; De Lima Araújo, E.; El-Deir, A.C.A.; De Lima, A.L.A.; Souto, A.; Bezerra, B.M.; Ferraz, E.M.N.; Maria Xavier Freire, E.; Sampaio, E.V.D.S.B.; Las-Casas, F.M.G.; et al. Caatinga Revisited: Ecology and Conservation of an Important Seasonal Dry Forest. Sci. World J. 2012, 2012, 205182. [Google Scholar] [CrossRef]
Alvares, C.A.; Stape, J.L.; Sentelhas, P.C.; De Moraes Gonçalves, J.L.; Sparovek, G. Köppen’s Climate Classification Map for Brazil. Meteorol. Z. 2013, 22, 711–728. [Google Scholar] [CrossRef]
Alves, E.D.L. Climatologia: Noções Básicas e Climas do Brasil. Soc. Nat. 2010, 22, 639–640. [Google Scholar] [CrossRef]
Andrade-Lima, D. The Caatingas Dominium. Rev. Bras. 1981, 4, 149–153. [Google Scholar]
IBGE. Manual Técnico da Vegetação Brasileira; IBGE: Rio de Janeiro, Brazil, 2012. [Google Scholar]
da Silva, J.M.C.; Leal, I.R.; Tabarelli, M. Caatinga: The Largest Tropical Dry Forest Region in South America; Springer: Cham, Switzerland, 2018. [Google Scholar]
Leal, I.R.; Da Silva, J.M.C.; Tabarelli, M.; Lacher, T.E. Changing the Course of Biodiversity Conservation in the Caatinga of Northeastern Brazil. Conserv. Biol. 2005, 19, 701–706. [Google Scholar] [CrossRef]
Sampaio, E.V.S.B. Overview of the Brazilian Caatinga. In Seasonally Dry Tropical Forests; Cambridge University Press: Cambridge, UK, 2010. [Google Scholar]
Santos, J.C.; Leal, I.R.; Almeida-Cortez, J.S.; Fernandes, G.W.; Tabarelli, M. Caatinga: The Scientific Negligence Experienced by a Dry Tropical Forest. Trop. Conserv. Sci. 2011, 4, 276–286. [Google Scholar] [CrossRef]
Key, C.H.; Benson, N.C. Landscape Assessment: Remote Sensing of Severity, the Normalized Burn Ratio and Ground Measure of Severity, the Composite Burn Index. In FIREMON: Fire Effects Monitoring and Inventory System; USDA Forest Service, Rocky Mountain Research Station: Ogden, UT, USA, 2005. [Google Scholar]
Pérez, C.C.; Olthoff, A.E.; Hernández-Trejo, H.; Rullán-Silva, C.D. Evaluating the Best Spectral Indices for Burned Areas in the Tropical Pantanos de Centla Biosphere Reserve, Southeastern Mexico. Remote Sens. Appl. 2022, 25, 100664. [Google Scholar] [CrossRef]
de Queiroz, L.P.; Cardoso, D.; Fernandes, M.F.; Moro, M.F. Diversity and Evolution of Flowering Plants of the Caatinga Domain. In Caatinga: The Largest Tropical Dry Forest Region in South America; Springer: Cham, Switzerland, 2018. [Google Scholar]
Pinto, O. An Overview of Cloud-to-Ground Lightning Research in Brazil in the Last Two Decades. In Atmospheric Science Research Progress; Nova Science Publishers Inc.: New York, NY, USA, 2009. [Google Scholar]
Melo, F.P.L. The Socio-Ecology of the Caatinga: Understanding How Natural Resource Use Shapes an Ecosystem. In Caatinga: The Largest Tropical Dry Forest Region in South America; Springer International Publishing AG: Dordrecht, The Netherlands, 2018. [Google Scholar]
Tabarelli, M.; Leal, I.R.; Scarano, F.R.; Silva, J.M.C. da Caatinga: Legado, Trajetória e Desafios Rumo à Sustentabilidade. Cienc. Cult. 2018, 70, 25–29. [Google Scholar] [CrossRef]
Argibay, D.S.; Sparacino, J.; Espindola, G.M. A Long-Term Assessment of Fire Regimes in a Brazilian Ecotone between Seasonally Dry Tropical Forests and Savannah. Ecol. Indic. 2020, 113, 106151. [Google Scholar] [CrossRef]
Abreu, M.C.; Lyra, G.B.; de Oliveira-Júnior, J.F.; Souza, A.; Pobočíková, I.; de Souza Fraga, M.; Abreu, R.C.R. Temporal and Spatial Patterns of Fire Activity in Three Biomes of Brazil. Sci. Total Environ. 2022, 844, 157138. [Google Scholar] [CrossRef]

Figure 1. Map of the boundaries of the Caatinga biome.

Figure 2. Overview of the method for classifying burned areas in Caatinga.

Figure 3. The Multi-Layer Perceptron Network‘s structure involves using the spectral bands (RED, NIR, SWIR1, and SWIR2) as input layers and the classes burned and unburned as the output layers.

Figure 4. The Multi-Layer Perceptron Network‘s structure involves using the spectral bands (RED, NIR, SWIR1, and SWIR2) as input layers and the classes burned and unburned as the output layers. (A) depicts the cumulative burn area from 1985 to 2023. (B) in contrast, showcases the annual burn area over the same temporal range.

Figure 5. The annual distribution of annual burned class areas in the Caatinga biome from 1985 to 2023.

Figure 6. The annual distribution of burned areas by land use and land cover types in the Caatinga biome from 1985 to 2023.

Figure 7. The paper presents the spatial distribution of fire frequency in Brazil from 1985 to 2023, including the corresponding burned area and proportion by frequency class. (A) shows the map of fire frequency throughout the Caatinga biome, while (B) presents the classes of fire frequency by area and their corresponding percentages.

Figure 8. The figures depict the spatial association between accumulated burn scars and various climate parameters. (A) illustrates the correlation between burn scars and accumulated precipitation. (B) showcases the relationship between accumulated burn scars and climate water deficit. Lastly, (C) presents the correlation between burn scars and reference evapotranspiration.

Table 1. Monthly areas affected by fire scars in the Caatinga biome (hectares) from 1985 to 2023. Month indicates the month of the year. Area (ha) is the total area affected by fire scars. % shows the percentage of total area affected.

Month	Area (ha)	%
January	897,161	4.78%
February	211,880	1.13%
March	118,808	0.63%
April	51,178	0.27%
May	83,345	0.44%
June	77,631	0.41%
July	240,352	1.28%
August	1,043,449	5.56%
September	3,418,562	18.20%
October	7,266,083	38.68%
November	3,707,726	19.74%
December	1,666,578	8.87%

Table 2. The table presents accuracy metrics used to evaluate the classification of mapped burn areas.

Year	Overall Accuracy	Precision	Re Call	F1 Score	Jaccard
2007	0.94	0.97	0.87	0.91	0.84
2013	0.98	0.98	0.95	0.96	0.93
2014	0.97	0.82	0.85	0.83	0.74
2015	0.92	0.94	0.85	0.88	0.80
2017	0.96	0.97	0.93	0.95	0.91
2018	0.97	0.95	0.85	0.89	0.82
2019	0.98	0.98	0.96	0.97	0.95
2020	0.96	0.97	0.87	0.91	0.85
2021	0.80	0.79	0.85	0.79	0.66
2023	0.95	0.94	0.96	0.95	0.90

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

© 2024 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).

Share and Cite

MDPI and ACS Style

Franca Rocha, W.J.S.; Vasconcelos, R.N.; Duverger, S.G.; Costa, D.P.; Santos, N.A.; Franca Rocha, R.O.; de Santana, M.M.M.; Alencar, A.A.C.; Arruda, V.L.S.; Silva, W.V.d.; et al. Mapping Burned Area in the Caatinga Biome: Employing Deep Learning Techniques. Fire 2024, 7, 437. https://doi.org/10.3390/fire7120437

AMA Style

Franca Rocha WJS, Vasconcelos RN, Duverger SG, Costa DP, Santos NA, Franca Rocha RO, de Santana MMM, Alencar AAC, Arruda VLS, Silva WVd, et al. Mapping Burned Area in the Caatinga Biome: Employing Deep Learning Techniques. Fire. 2024; 7(12):437. https://doi.org/10.3390/fire7120437

Chicago/Turabian Style

Franca Rocha, Washington J. S., Rodrigo N. Vasconcelos, Soltan Galano Duverger, Diego P. Costa, Nerivaldo A. Santos, Rafael O. Franca Rocha, Mariana M. M. de Santana, Ane A. C. Alencar, Vera L. S. Arruda, Wallace Vieira da Silva, and et al. 2024. "Mapping Burned Area in the Caatinga Biome: Employing Deep Learning Techniques" Fire 7, no. 12: 437. https://doi.org/10.3390/fire7120437

APA Style

Franca Rocha, W. J. S., Vasconcelos, R. N., Duverger, S. G., Costa, D. P., Santos, N. A., Franca Rocha, R. O., de Santana, M. M. M., Alencar, A. A. C., Arruda, V. L. S., Silva, W. V. d., Ferreira-Ferreira, J., Oliveira, M., Barbosa, L. d. S., & Cordeiro, C. L. (2024). Mapping Burned Area in the Caatinga Biome: Employing Deep Learning Techniques. Fire, 7(12), 437. https://doi.org/10.3390/fire7120437

Article Menu

Mapping Burned Area in the Caatinga Biome: Employing Deep Learning Techniques

Abstract

1. Introduction

2. Materials and Methods

2.1. Study Area

2.2. General Strategy

2.3. Classification and Sample Regions

2.4. Deep Learning Architecture Model

2.5. Data Analysis

2.6. Validation Analysis

3. Results

3.1. General Patterns

3.2. Extent of Burned Areas

3.3. Monthly Fire Season Patterns

3.4. Burn Area in LULC

3.5. Fire Frequency and Spatial Patterns

3.6. Fire Dynamics in Subregions Within the Caatinga Biome

3.7. Accuracy Analysis

3.8. Correlation Between Fire Burn Scars and Climate Variables

4. Discussion

5. Conclusions

Supplementary Materials

Author Contributions

Funding

Institutional Review Board Statement

Informed Consent Statement

Data Availability Statement

Acknowledgments

Conflicts of Interest

References

Share and Cite

Article Metrics

Article Access Statistics

Further Information

Guidelines

MDPI Initiatives

Follow MDPI