DYNAMICS OF AGRICULTURAL EXPANSION IN AREAS OF THE BRAZILIAN SAVANNA BETWEEN 2000 AND 2019

The information survey and the land use and land cover (LULC) change monitoring are essential to understand the changes in the landscape and their impacts on the environment. The Brazilian savanna (Cerrado) constitutes the second largest biome in Brazil and is highly relevant because of its rich biodiversity. The Cerrado in the Maranhão State is facing a high conversion rate of its natural vegetation into agricultural systems because of the agricultural development policies. This article aims to analyze and quantify the LULCC in the Chapadinha microregion, Maranhão State, Brazil, by comparing Landsat satellite images from 2000, 2009, and 2019. The Chapadinha microregion covers an area of more than 14,000 km2 and, since 2000, presents significant spatial transformations related to agricultural expansion. The goal of this expansion is the economic development based on agricultural commodities (mainly soybean monoculture) for exportation. The study area is the third microregion in terms of agricultural production in Maranhão. The satellite image interpretation showed a reduction of more than 800 km2 of natural vegetation.


INTRODUCTION
Since 1970s, the Cerrado biome has consolidated as the main region of the country in terms of expansion of the agricultural because of the favorable climate and soil, and the large amount of land available for cultivation. As a result, Brazil is one of the largest grain producers and exports in the world, with emphasis on soy production. Cerrado occupies approximately 20% of the Brazilian territory (~ 2 million km 2 ) and is distributed in the states of Goiás, Tocantins, Mato Grosso, Mato Grosso do Sul, Minas Gerais, Bahia, Maranhão, Piauí, Rondônia, Paraná, São Paulo, and the Federal District (Sano et al., 2009).
The combined influence of climate (highly seasonal), topography (flat or steep, depending on the location), and soils (highly weathered, deep, and acidic) controls the formation of different grass-, shrub-, and tree-dominated phytophysiognomies with varying proportions of these three strata. Cerrado has more than 12 thousand cataloged plant species -approximately 40% are endemic (MMA, 2015), therefore, it is included as one of world´s hotspots for biodiversity conservation. In the same time, it is considered the last agricultural frontier in the country (Klink and Machado, 2005). Government incentives and private capital associated with the Cerrado Development Plan (Santos, 2011) increased the suppression of natural vegetation and the implementation of large-scale production of agricultural commodities (Inocêncio e Calaça, 2009).
In Maranhão, the Cerrado occurs from southern portion to the north, covering 64% of its territory (74,288 km²). Recent natural vegetation suppression in this state is associated with Prodecer III project that started in the Gerais de Balsas microregion in the 1990s. Soybean cultivation in the Maranhão started in 2000s, * Corresponding author when agricultural frontier began to shift from western Bahia State to the east of Maranhão. The emphasis was in the Chapadinha and Baixo Parnaíba microregions, considered as the Cerrado´s new agricultural frontier (Gaspar, 2010, Santos, 2011, Campos, 2011. In the 2000s, the study area presented the first agricultural modernization process with the insertion of the monoculture crop. Its consolidation started in 2003 with emphasis on the municipalities of Brejo and Buriti, since they present landscape characteristics favorable to the development of highly mechanized agriculture in flat terrains, regardless of dominant well-drained, low fertility Oxisols developed from pre-weathered sediments of the Barreiras Formation. In addition, the prices of lands are still relatively low (Guimarães, 2012).
The use of computer tools and remote sensing techniques to survey characteristics of the region of interest are highly relevant to define and implement biodiversity conservation strategies, reduction of environmental impacts, and environmentally sustainable agriculture planning (Bendini et al., 2016, Sano et al. 2019a. The Brazilian savanna is a key region in terms of expansion and consolidation of the national agricultural development and the applications of remote sensing for analysis of monitoring agriculture, land use and land cover is very important for the region (Arantes et al. 2016, Bolfe et al. 2016, Picoli et al., 2018, Araújo, et al. 2019, Camargo et al., 2019, Alencar et al., 2020. This paper aims to analyze the dynamics of land use change in the Chapadinha microregion, Maranhão State, by identifying and quantifying the land use dynamics during the 2000-2018 period. This study also relates the historical and socioeconomic activities with the processes of transformations and their impacts.

Study Area
The Chapadinha microregion occupies approximately 10,400 km² in the northeast of Maranhão State. It encompasses the municipalities of Anapurus, Belágua, Buriti, Brejo, Chapadinha, Mata Roma, Milagres do Maranhão, São Benedito do Rio Preto, and Urbano Santos (Figure 1). Regardless of presence of high-tech, mechanized agriculture that represents more than 80% of the current production model in the area, the majority of population still survive based on extractivism, agriculture practiced in itinerant way (cutting, slashing, and burning), low-productivity cattle ranching and fishing.
The study area is located in a transition region between the Cerrado and Caatinga biomes (Batistella et al., 2013). There is predominance of arboreal forest formations belonging to the Cerrado domain, mixed forest, and extensive fields of open vegetation in tabular zones. The microregion presents high temperatures and annual rainfall ranging between 800 mm and 1200 mm, and humidity ranging between 70% and 80%. It presents two distinct climate periods, a rainy season and a dry season, differentiated landscapes with high vegetation variety, depending on the seasonality (Reschke, 2013).
The initial human occupation in the study area is strongly associated with the extensive cattle ranching. Since 1990s, the Chapadinha microregion integrates the state's agricultural production chain, with large-scale soy production (Gaspar, 2010). According to Almeida (2017), the process is directly associated with the improvement of plant and soil management technologies. In fact, this improvement boosted local production and productivity of soybean for exportation and agroforestry for cellulose, causing large suppression of natural vegetation cover.
According to data provided by the Municipal Agricultural Production (PAM) (IBGE, 2019), soybean is responsible for 85% of the total area of croplands. Other crops found in the study area are rice, corn, and cassava. Data from IBGE showed that, there were 411 hectares of soybean in 2000, increasing to 65,240 hectares in 2018. The intensification of agricultural production is observed since 2002, when the area destined for soybean cultivation increased from 550 hectares to 3,716 hectares in 2003. The production increased from 1,320 tons to more than 8,900 tons ( Figure 2).

Images
The LUCL mapping was carried out based on visual interpretation of a set of Landsat satellite images from 2000 to 2019. The images were downloaded from the United States Geological Survey Earth Explorer Platform (USGS, 2020). The selected images were from August to December, with less than 5% of cloud cover (Table 1). The following digital image processing tasks were performed: i) registration, radiometric transformations, and visual adjustments, ii) generation of the normalized difference vegetation index (NDVI), iii) image segmentation and extraction of attributes, iv) data mining and image classification based on the decision tree algorithm, and v) analysis of the results.

Image Processing
To minimize the atmospheric influence, the images were preprocessed. Initially, the gray level digital numbers were converted into radiance and then to the surface reflectance following the methodology described by Ponzoni and Shimabukuro (2009) and using the Spring 5.6 software. For the The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLIII-B3-2020, 2020 XXIV ISPRS Congress (2020 edition) This contribution has been peer-reviewed. https://doi.org/10.5194/isprs-archives-XLIII-B3-2020-1607-2020 | © Authors 2020. CC BY 4.0 License. conversion of Landsat-5 images, we used the equations described by Rosa (2009)  The radiance at the top of the atmosphere was calculated from the function described by Ponzoni and Shimabukuro (2009) where dr = inverse of the square of the Earth-Sun relative distance in astronomical units z = zenital solar angle (degrees) at the time of image acquisition Eλ = average solar irradiance at the top of the atmosphere (mW/cm².Ω.µm) Lλ = monochromatic spectral radiance (W/m².sr.µm) ρλ = monochromatic reflectance.
For the calculation of radiance and reflectance of the Landsat-8 images, we observed changes in the conversion parameters, showing a multiplicative factor and an additive factor calculated separately and following Eq. 3: where ρλ '= spectral reflectance Qcal = image to be transformed M = multiplicative factor for the reflectance of the band to be converted A = additive factor for the reflectance of the band to be converted.
The images had the process of atmospheric correction to minimize the effects of atmospheric scattering. The dark-object subtraction technique of multispectral data -DOS (Chaves, 1988) was applied using the QGIS software through the Semi-Automatic Classification Plugin (Congedo, 2016).
Based on radiance and reflectance values, the NDVI was obtained as a tool for monitoring plant vigor, since NDVI is sensitive to green biomass variation, chlorophyll content, and water stress (Sano et al., 2019b). Its calculation based on the relationship between the red and near infrared bands (4).
where NDVI = normalized difference vegetation index PNIR = near infrared band PR = red band 1 Similarity threshold / area 10/100 was established for Landsat 5, and for Landsat 8 50/300, following the parameters described in Meneses and Almeida (2012).
After the conversion, a histogram contrast was applied. It consists of the pixel scattering process with a linear function of 1 being applied, without changes in the original spectral characteristics of the targets (Meneses and Almeida, 2012). The images were clipped to the boundaries of the Chapadinha microregion provided by IBGE (2019). We then generated a merged image with all multispectral bands and NDVI. The metafile was elaborated in the TerraView image visualization software.
The images were also segmented and classified by using the Geographic Data Mining Analyst (GeoDMA) platform (Korting et al., 2009). It integrates methods of image analysis with data mining techniques from spatial and spectral information, functioning as a TerraView plug-in (Korting et al., 2009).

Segmentation, Attribute Extraction, and Sample Collection
Image segmentation consists of a process of grouping pixels having similar characteristics in discrete and contiguous regions without intersecting themselves and constituting in segments (objects). We used segmentation by growing regions (Baatz and Schape, 2000), which, according to Benz et al. (2004), is a technique of pixel aggregation that starts with a pixel (seed), predefined values of similarity 1 , threshold, and scale, grouping neighbors having similar properties and considering the spectral and spatial characteristics in the composition of objects (Zanotta et al., 2019).
After the segmentation process, attributes related to the relational objects present in the image and associated with spatial (13) and spectral (15) characteristics were extracted, resulting in a total of 28 attributes for each scene or 112 attributes for each image. This step was performed using the Geodma plug-in available in the TerraView 5.3.3 software.
Six LULC classes were defined: Field, Forest, Bare Soil, Agricultural Areas, Water, Built Area. At least 68 sampling points were collected for each LULC class, according to the methodology proposed by Sano et al. (2009).

Classification and Validation
This step was performed by Geodma. We used the decision tree procedure (Korting et al., 2009)  The process starts with the definition of hierarchy of data by internal nodes and leaves connected by branches. Each node corresponds to a variable that is used for classification. The first node is known as the root, and the others are called intermediate. The leaves correspond to the variables related to the classes that each intermediate node is associated (Korting et al., 2009).
Initially, three decision trees were created, aiming to identify trees that presented a lower branching pattern and better performance, allowing their use for all images. For this purpose, the Kappa and the performance indices were used as the validating model to classify the maps generated. The maps were analyzed and compared with the thresholds established by Landis and Koch (1977).
where k = index value n = total number of samples Σdp = sum of the main diagonal Σ(l*c)= product of the sum of the row by the columns of each representative class.
To validate the classifications, a cross combination was carried out with ground truth data and GeoEye images that are available for free in the Google Earth platform.

Change Detection
After correcting the errors associated with the classification, the vector data related to land cover classes were converted into matrix information using conversion tools available in the QGIS 2.18.8 software. The purpose here was to perform change detection analysis and to generate transition matrices among the mapped classes. Thus, the Molusce (Modules for Land Use Change Simulations) alteration detection statistics available in the QGIS were applied. The results generated tables with information on the class area and frequency of coverage, besides reference matrices related to the years 2000 to 2009, and to the years 2009 to 2019. The results identified a percentage of the loss and gain rates between the classes.

Validation of the Classification
LULC changes, as well as to identify the advance of deforestation in the Cerrado associated with the advance of the agricultural frontier in 2000, 2009, and 2019. The Landsat satellite images allowed the mapping of the main patterns of land cover changes. The validation of the classification models was performed using the confusion matrix and the Kappa index.
The overall accuracy rate were 76%, 67% and 59% respectively for the years 2000, 2009 and 2019. The forest class hit were 77%, 87%, and 72%, and the agricultural areas were 87%, 88%, and 88% for the classified classes. It is worthwhile to mention the low hit rate for the silviculture (17%, 8%, and 28%), with a high rate of confusion with the forest and agricultural areas. This relatively high rate of confusion is possibly associated with climatic of drought during the acquisition of the analyzed images and the natural conditions of cerrado, as expected in the naturally heterogeneous cerrado (Alencar et al., 2020).

Land Use and Land Cover
The generation of thematic LULC maps allowed the mapping and quantifying transition rates, in addition to the analysis of agricultural expansion in the area. We can observe the temporal advance of agriculture over natural areas (grasslands and forests), mainly in the eastern portion (municipalities of Brejo, Buriti, and Anapurus), and silviculture in the northern portion (Urbano Santos municipality). The main changes observed in the area are related to agricultural expansion over the areas of natural vegetation, with a gradual reduction in the areas of forest and grasses. In 2000, these classes represented 6617.4 8km² and 3523.81 km², respectively. In 2019, they were reduced to 5755.28 km² and 3388.43km², that is, a reduction of approximately 863 km² of forested areas, and 135 km² of grass-dominated areas.
Along the period, a significant increase occurred in new agricultural and forestry areas in the Chapadinha microregion. The data showed an increase of more than 800% between in the 2000-2019 period. This was boosted mainly by state policies in association with private capital, as highlighted by Gaspar (2010) and Almeida (2017). The pattern of agricultural expansion was defined as selective, predominantly in the east and center of the microregion, mainly in the municipalities of Brejo, Buriti, Anapurus, and Mata Roma.
The predominance is associated with geomorphological characteristics and the topographic profile (composition of tabular features with slope varying from flat to smooth) that enabled the implementation of modern agricultural practices and the use of fertilizers, resulting in significant agricultural advances. Araújo et al. (2019) emphasized that the municipality of Buriti presented 78% of new areas destined to agricultural production between 2000 and 2007. Almeida (2017) noted that the municipality of Brejo presented significant increase in areas for agricultural production between 2000 to 2015. When analyzing the vegetation cover conditions in the region,  highlight the area has shown high rates of deforestation in recent decades.

Detection of Change -2000 to 2009
Table 6 describes the areas occupied by each LULC classes, and the corresponding percentages of transitions that occurred in Chapadinha microregion between 2000 and 2009. The forest class showed significant reduction, approximately 7.13%, or a total of 473 km². The main conversions from forest areas were to fields, approximately 280 km², agricultural areas 133 km², and forestry 22 km². The regeneration areas totalized approximately 773km², 21% of which came from rural areas, 13% from exposed soil areas, and 38% expansion. It is worth noting that the class had a maintenance rate of 81%, which indicates 5517.78km² of forest areas maintained between 2000 and 2009.   Table 6). The main gains are associated with the conversion of rural areas (~453 km²), and areas classified as bare soils (211 km²). The conversion of agricultural areas into other classes was low (<10%). Figure 6 shows the transition values between the LULC classes. Significant advance of agricultural areas and reduction of forest and grasslands was observed.

Change Detection -2009 to 2019
Between 2009 and 2019, the area faced a gradual change in terms of LULC classes (

CONCLUSIONS
The use of Landsat images was essential to support the interpretation of the different types of LULC classes in the study area. It enabled the creation of an updated base of conversion and advance of soybean production in the area. The use of digital image processing techniques proved to be satisfactory to assess and analyze changes and to identify temporal changes at municipality level. The data survey of images and classifications and their insertion in a geographic database generated information about LULC changes. The object-oriented classification method was considered significant, as it allowed great automation of the mapping process and reduction in the efforts of manual editing's.
There was a 13% deforestation rate in the analyzed period, with high percentage associated with the insertion of annual crops. The agricultural area increased over 500 km² between 2000 to 2019. Despite the high rate of deforestation, forest and countryside areas still represent the predominant land cover classes in the microregion. It represents approximately 88% of the land cover, mainly in the western and southern parts of the microregion, with emphasis to the municipalities of Chapadinha and Belágua that had a deforestation rate below 20%. The municipalities of Brejo, Buriti, Anapurus, and Mata Roma presented deforestation rates higher than 45%.
The data may serve as an instrument to elaborate protective environmental policies for the Cerrado at municipality level, as well as a management tool for public and private sectors.