This study was conducted in the conservation area denominate Parque Estadual do Tainhas (PET) and its buffer zone (BZ) located at latitude 29°5′15″S and longitude 50°22′4″W (Figure 1). The PET was established in 1975 to protect the grasslands and forests in the Tainhas river valley, inserted in the AFB.

The Park is formed by mosaics of forests with Araucaria pine, grasslands, and flooded areas (wetlands) with flatter terrain in the southern portion to valleys in the northern part (SEMA, 2008). It has an area of 6,654 ha with a BZ of approximately 60,000 ha, covering the municipalities of Jaquirana (69.8%), São Francisco de Paula (20.6%), and Cambará do Sul (9.6%) of the Park’s area. Around the PET, communities in its BZ depend on agriculture, forestry, and livestock and regularly use fire as a management tool (Bond-Buckup, 2008).

The PET is close to other protected areas, which favors the emergence of the biogeographic conditions that can contribute to the interconnection of these units through ecological corridors, helping preserve fauna and flora populations and maintain the ecosystem services balance in the region.

The methodological approach for burned area detection used in this study is shown in the flowchart (Figure 2). We use Sentinel-2 MSI Multispectral Instrument, Level-2A data. The Level-2A product provides Bottom Of Atmosphere (BOA) imagery derived from the associated Level-1C products. The imagery used is available from the Google Earth Engine (GEE). This platform combines three tools (code editor, explorer, and client libraries) for users to perform geospatial analysis in the cloud (Adagbasa, Adelabu, and Okello, 2020). We select images from 2018 to 2022, before the start of the fire season, from January 1 to April 1, denominating as “pre-fire” images, while the post-fire images were selected from July 15 to September 15 of each year, when the “burning window” allowed by municipal legislation was established. A cloud mask was applied. First, the function defines two bitmask values for clouds and cirrus (bit 10 and bit 11, respectively) and selects the image’s pixel quality band (QA). Then, the function creates a mask that filters out all the pixels where the cloud and cirrus bits are equal to zero, indicating clear conditions.

We used the GEE to collect spectral signatures of the burned and unburned areas, which served as samples for the model classification. In addition, we used data collected from the field in the year 2020 as a reference for sampling in this year’s post-fire images. This effort resulted in 7,133 sampled pixels, manually collected as small polygons from burned areas (2,295 sampled pixels) and the unburned regions (4,838 sampled pixels).

For training, we used the shortwave infrared (SWIR) and near-infrared (NIR) bands and the result of the Normalized Burn Ratio spectral index (ΔNBR) calculation.

The NBR is a ratio of the NIR to the SWIR region, developed to identify post-fire burned areas and provide a quantitative measure of burn severity (Key and Benson, 2006). The NBR is calculated by the pre- and post-fire difference (denoted as ΔNBR) using Eqs. 1, 2. N B R = N I R − S W I R N I R + S W I R (1) Δ N B R = N B R p r e − f i r e − N B R p o s − f i r e (2)

The algorithm used for our classification was the Random Forest (RF) (Breiman, 2001; Goehry et al., 2021), which is an ensemble algorithm operated by building multiple decision trees in a training session and assigning the target class by majority vote (PAL, 2005). In this study, we used the RandomForest function in the GEE library. The parameters used were: number0fTrees (20), variablesPerSplit (null), minLeafPopulation (1), bagFraction (0.5), maxNodes (null), and seed (0). After classification, we applied a spatial filter to remove noise and fill in gaps, where burned areas smaller than or equal to 1 ha (5 × 5 pixels) were removed, and, in the same way, internal spaces were filled as burned. The spatial filter selected was the Manhattan Kernel, which generates a distance kernel based on rectilinear distance (city-block). Reduction is performed by calculating the mode (most common value) of the pixel values in a neighborhood defined by the specified kernel (or window). The filter size was determined considering the difference in the spatial resolution of the Landsat 8 satellite, which was used for the MAPBIOMAS product methodology proposed for exclusion. Finally, the Receiver Operating Characteristic (ROC curve) was calculated to measure and compare the binary classification model’s performance.

To evaluate the management tools available in conjunction with the classification data generated, we used data on the annual area burned by MapBiomas Fire—Collection 1.0 (Alencar et al., 2022) and the land use and land cover areas classified as grassland from 1985 to 2020 available in MapBiomas Collection 7.0 (Souza et al., 2020). In addition, we also used the extent area authorized for burning by Jaquirana, São Francisco de Paula, and Cambará do Sul municipalities located within PET and its BZ.

The annual burned areas stationarity trend covered by grassland was checked using the Augmented Dickey-Fuller (ADF) test. The ADF test is an “augmented” version of the Dicker-Fuller test. The ADF test expands the test equation to include high-order regression processes in the model.

Afterward, the trend was verified through time series decomposition to verify the seasonality and the residuals. Results are obtained by first estimating the trend by applying a convolution filter to the data. Therefore, the trend is removed from the series, and the average of this unbiased series for each period is the seasonal component returned (McKinney, Perktold and Seabold, 2011).

The burnt areas and the grasslands from 1985 to 2020 in the PET and its BZ are shown in (Figure 3). In this figure is highlighted in vertical dotted lines, the primary legislation, which is also presented in (Table 1) together with the grasslands and the total burned area values. This analysis reveals an increasing trend for the burned area, while there is a decreasing trend for grassland in the PET.

We can observe that the burnt areas remained low after the fire use prohibition in 1992. However, this pattern was not established for a long time, and the highest values for the time series analyzed occurred in 1995 and 1996, with about ten thousand hectares burnt each year.

The most significant burned areas reduction occurred in 2002, reaching the lowest value in 2005, with only 35 ha. By observing the trend curves of burnt area and area occupied by grassland formation, the curve’s decline in the same years is noticed.

In 2002, the area corresponding to grassland was 49,359 ha, while in 2005, it was 46,345 ha. Therefore, the PET and its BZ lost more than 3,000 ha in only 3 years. These years precede the AFB Law creation (Brasil, 2006) and its regulation decree (Brasil, 2008), prohibiting new conversions of areas with native vegetation in the biome.

During the 35 years of the analyzed time series, 12.285 ha of grassland areas were converted to other uses, losing about 24% of their original formation, with 10% occurring after the AFB decree banned this practice in 2008 (Brasil, 2008).

Finally, concerning the burned areas only when municipalities begin to legislate about the use of fire, in 2013, a trend definition and a biannual frequency can be observed, as provided by several municipal legislation (Municipio de Cambará do Sul, 2013; Municipio de Jaquirana, 2013; Municipio de São Francisco de Paula, 2013).

The burned areas classification with the RF algorithm obtained an AUC = 0.9983 (Figure 4). With a visual inspection, the scar’s demarcation can be identified when comparing the pre-fire and post-fire images (Figure 5).

The results show that larger areas were burned for all years than those allowed by the municipalities’ environmental permits. In addition, larger burned areas were found in our classification than those presented by the MapBiomas Fire collection, as can be seen in (Table 2).

The licenses issued correspond to only 48.09% (2018), 33.62% (2019), 23.55% (2020), 69.43% (2021), and 43.54% (2022) of the area identified as burned in our classification. For the three available years of the analyzed period from the MapBiomas collection, they represent only 26.34% (2018), 29.69% (2019), and 49.74% (2020) of the burned areas.

The Mapbiomas burned areas that coincide with our classification represent only 21% of the total classified area on average (Figure 6). The MapBiomas burned omission areas are formed by scars smaller than 10 ha or with sites that present a non-continuous formation with gaps of 1 ha.

In addition to the burning extent, the frequency of areas affected by the fire was verified, and about 28% of the PET and its BZ were burned at least once, as shown in (Figure 7). Notably, according to the legislation, the areas could present a burning frequency of 2–3 times for 5 years. However, the results showed that 353 ha were burned 4 to 5 times, exceeding the legal requirements for this period. These areas are within the Jaquirana municipality territory inside the PET or near its boundaries.

No stationary trends were found in the time series, even though there is a long history of legislation to regulate fire use and vegetation protection practices.

During the 20 years (1992–2012) of fire use practice prohibition, producers changed their economic profile, and extensive native grassland areas conversion to monocultures in the region occurred (Boziki, Beroldt, Printes, 2011; Buffon, Printes, Andrades-Filho, 2018). Grasslands are important for livestock, especially where other agricultural practices are not viable. In this study, we noticed that legislation changes over the time series are related to changes in land use.

Conservation areas (CAs) in Brazil are managed under the National System of Conservation Units (SNUC), which is designed to protect the country’s diverse ecosystems (BRASIL, 2000). The SNUC comprises 12 categories of CAs, each with specific objectives regarding protection and use. The PET falls under the Park category, which provides full protection and allows only non-consumptive uses of natural resources. This means that handling fire is not permitted inside the PET. Conservation areas like the PET are essential for preserving Brazil’s rich biodiversity and ensuring the long-term sustainability of the country’s natural resources.

The fire management practice in the winter period is expected in the region, and it is used to remove dry biomass to provide the vegetation regrowth that will be used to feed the cattle herd in spring and summer (Pillar et al., 2009).

Once forbidden, it can be replaced by other activities much more environmentally damaging than the old fire management grasslands practice, such as forestry and monocultures, which have been advancing into the grasslands (Buffon, Printes and Andrades-Filho, 2018).

It was observed during the study that municipalities lack data and tools to manage and enforce current environmental legislation. So it is necessary to evaluate the effectiveness of existing environmental management tools regulated by municipal laws related to the use of fire (Santos and Andrades-Filho, 2021).

In addition to the environmental laws presented, it was verified the irregularity in the data availability by the municipalities, which do not meet the legal provisions of free access to information in Brazil, especially regarding publicity, accessibility, and transparency, making technological advancement impossible (Brasil, 2011).

The brief characteristics of the fire marks left on satellite imagery complicate the burned area’s detection. There are few fire products available globally, and only one at the national level (Alencar et al., 2022). So, evaluating its applicability to different regions and vegetation formations in Brazil is extremely important.

Our study found significant differences in areas mapped as burned, increasing by more than 50% in our study using a better spatial resolution. The improvement of available products is substantial and progresses with the recent technologies emergence. In previous studies of Russian grassland, the difference in the area mapped between satellites with lower spatial resolutions also increased dramatically from no burning to as much as 19% of the total study area (Dubinin et al., 2010). The strategy of using ΔNBR in conjunction with the NIR and SWIR bands for burned areas classification reduced noise and class confusion, especially with wetlands. The areas with the highest commission error were exposed soil from recently cleared forestry areas.

Due to the rugged relief and mosaics of forest and mountainous grassland vegetation that form the area’s landscape, remote sensing data with higher spatial resolution can be explored as they become available. Therefore, Earth observation can significantly support public fire prescription policies and add to other factors that consider CO² emissions (Herrmann, Nascimento and Freitas, 2022).