1. Introduction
Humans historically have dominated their geographical surroundings, interacting, and modifying it according to their interests. Unfortunately, this often causes degradation of the natural resources due to the different forms of environmental impacts [
1]. An important example involves changes in land cover, which, when associated with the absence of conservation practices, generate impacts such as a reduction in water supply [
2,
3] resulting from changes in the hydrological cycle [
4,
5]. In Brazil, this has involved deforestation of the Amazon rainforest and habitat conversion of the Cerrado savannahs for agricultural use [
6,
7]. Natural habitat areas are typically converted to pasture for extensive livestock (e.g.,
Bos taurus Nelore beef breed) grazing and/or the export of commodity crops, such as soybeans (
Glycine max L.), maize (
Zea mays L.), and cotton (
Gossypium sp.). These recent land use conversions and the resulting agricultural economic development have been driven by international demand for food and livestock feed [
6]. Continued habitat conversions and/or changes in the climate could make the sole reliance on rain-fed agriculture more challenging if precipitation continues to decline in the Amazon [
8].
In order to meet the growing needs of humans (e.g., food) and mitigate the environmental impacts of agricultural production by more efficiently using natural resources, it is essential to adopt planning policies that integrate environmental, social, and economic aspects [
9]. Monitoring changes in land classifications (i.e., classes), as a function of different land uses, has become an effective tool to support land use planning. In particular, the use of technologies such as remote sensing, which can currently count on a variety of sensors operating at different scales, allow for the acquisition of information on land use classifications over large areas at low cost [
10,
11].
Data resulting from remote sensing can be easily processed using geoprocessing programs that make it possible to carry out different types of operations to generate information. In studies that aim to characterize land use and occupation, the application of the image classification method [
12] stands out. This method consists of labeling the image pixels according to their spectral characteristics, using, for this purpose, mathematical techniques that perform the pattern recognition resulting in thematic maps [
13].
Due to the global importance and recognition for the ecological services it offers, the Amazon has been monitored for decades using remote sensing and image classification products. This monitoring has sought to evaluate changes to this rich biome in order to help implement public policies for Amazon conservation. In a study carried out in the Colombian Amazon, Landsat images and supervised classification were used to map the changes in land cover and identify locations affected by deforestation over a period of sixteen years, between 2000 and 2016 [
14]. Other researchers have dedicated themselves to mapping and monitoring forest changes in the Brazilian Amazon in the state of Pará from 2000 to 2019, using multitemporal remote sensing data and machine learning classification [
15]. In the upper Teles Pires River basin’s transition zone between the Amazon and Cerrado biomes, previous research mapped the spatial and temporal dynamics of land use from 1986 to 2014, using Landsat images and supervised classification. The results showed an intense reduction in native vegetation as a result of agricultural expansion [
16].
Several other studies around the world have already employed remote sensing to identify the forms of appropriation of spaces and changes in landscapes. Changes in land use in poor areas of China were mapped between 2013 and 2018 [
17]. Another study monitored land cover changes in a district of India between 1990 and 2010 [
18]. Other studies focus on mapping specific targets, such as that conducted by researchers [
19], who monitored the urban spatial–temporal dynamics in Nagpur, India, between 1991 and 2010. In Brazil, one of the nationwide actions for mapping land cover and land use is the MapBiomas project, aimed at the conservation of the different Brazilian biomes, which has generated a historical series of annual maps from an initiative involving a collaborative network of specialists [
20].
Our present study focused on mapping the changes that occurred in the Teles Pires River basin (
Figure 1). The largest part of the Teles Pires River basin is located in the state of Mato Grosso, Brazil. The Mato Grosso state is characterized by great socioeconomic and ecological diversity, where the Pantanal, Amazon and Cerrado biomes share space [
21]. Both the Amazon and Cerrado biomes are found within the Teles Pires River basin. The region has been in full economic development mode, driven by industrial agricultural exports. In the last decade, Mato Grosso has installed large hydroelectric projects, which has resulted in profound changes to the landscape, pointing to the need to monitor such changes in view of the possible impacts on the environment. Despite this region’s importance in the national and international context, the region lacks continuous monitoring of the changes in land use resulting from agricultural expansion by large farming enterprises. The goals of our research were to expand the knowledge about the land use dynamics in the region in order to better manage the water resources and plan economic activities. The objective of our study was to evaluate the changes in land use in the Teles Pires River basin over a 34-year period from 1986 to 2020.
2. Using Remotely Sensed Data for Conservation
Despite the current abundance of research that generates information on land occupation and use, globally, there is still a lack of adequate information that allows for assessing the intensity of the changes that occur in terms of land use. This makes it impossible to estimate and evaluate the effects, impacts, or potential expansion of agricultural production in ecologically diverse biomes [
22]. Gradually, this deficiency has been overcome thanks to the use and improvement of technologies such as remote sensing, which facilitates the acquisition of information in large areas and allows for the collection of historical data [
23].
The longest record of orbital images of the Earth’s surface already covers five decades and corresponds to the Landsat mission, which is a partnership between the National Aeronautics and Space Administration (NASA) and the United States Geological Survey (USGS), having launched the first satellite in 1972, the Landsat 1 [
24]. Over the years, the Landsat mission has outdone itself with the launch of satellites carrying increasingly innovative sensors. Currently, the Landsat series includes the Landsat 8 and 9 satellites, the last one having launched in September 2021 [
25]. The Landsat data set currently provides a global basis for monitoring the changes in environments due to the expansion of human occupation, data that are freely available to the public and open source [
24].
Through the use of software and geoprocessing tools, the data obtained through remote sensing, such as satellite images, can be treated and processed, allowing for the extraction of information of interest and the generation of products, such as thematic maps. One of the most used processes in the generation of thematic maps from remote sensing images is so-called image classification, which can also be distinguished as being either supervised or unsupervised. The unsupervised classification is characterized by not requiring prior knowledge of the study area, as the algorithm examines the unknown pixels of the image and divides them into different classes. Meanwhile, the supervised classification demands knowledge of the study area, as it is up to the analyst to select the training pixels representative of each class, so that the algorithm performs the classification [
26]. One of the most popular supervised classification methods used is the maximum likelihood algorithm [
18].
The maximum likelihood algorithm is based on the probability that each pixel in the image belongs to each of the classes identified by the analyst during training. The maximum likelihood algorithm then assigns pixels to the class with the highest likelihood of being in a particular classification [
27]. For this, it evaluates the variance and covariance of the spectral response of the training class when classifying the unknown pixels, generating accurate results, as it is based on statistical parameters [
28].
With rapid technological advancement occurring in this area, several new methodologies have been developed and applied in satellite image processing [
29], such as machine learning techniques [
30], the random forest algorithm [
31], neural networks, and others [
32]. Even so, classic methods such as maximum likelihood continue to be used due to their easy access and availability in various software, and when well executed they result in land use and occupation data of satisfactory accuracy [
33].
In order to measure the quality of the information generated in the classification of images, it is essential to validate the results, identifying the accuracy of the mapping. This can be easily obtained by comparing a set of classified pixels with terrestrial truth data [
12]. The most common way of representing the accuracy of the classification of remote sensing data has been the use of an error matrix, which is the basis for a series of statistical analyzes, such as general accuracy and the kappa index [
34]. Together, the error matrix and the kappa index have come to represent the standard way of evaluating the accuracy of image classification [
12]. Following such verification to improve data accuracy, such remotely sensed data become more reliable for use. Thus, maps resulting from image classification can be used to follow-up and monitor changes in land use and occupation, especially in areas threatened by environmental degradation, such as tropical rainforests and savannahs.
In tropical regions, monitoring human intervention in environments is essential, given the intense pace of the conversion of the natural areas into arable areas. About half of the world’s remaining tropical rainforests are in Latin America. The tropical rainforests in both Central America and South America are also experiencing the world’s highest rates of deforestation, largely driven by large-scale commercial agricultural production [
35].
Based on mapping global deforestation footprints between 2001 and 2015 [
36], tropical forests are under increasing threat. This is especially the case in tropical countries, such as Brazil, with high historical deforestation footprints that have allowed for the production and export of agricultural commodities, such as cattle, soybeans, coffee, cocoa, and wood, to other countries. Thus, it is possible to associate spatial patterns of deforestation with global supply chains [
37]. Massive investments have been made to support the production of export commodities in tropical countries, resulting in high rates of deforestation [
38]. This activity requires significant conversion of the land for use. Governments sometimes see these investments as beneficial, by improving the use of land seen as idle, disregarding that many of these lands are occupied by traditional peoples and ignoring the ecological importance of natural systems [
38].
4. Results
The Teles Pires River basin has an area of 141,524 square kilometers (km
2), of which 34,453 km
2 corresponds to the region of the upper Teles Pires, 55,890 km
2 to the middle Teles Pires, and 51,181 km
2 to the lower Teles Pires. These regions represent 24.34%, 39.49%, and 36.16% of the basin, respectively.
Figure 3 and
Figure A1 show the maps obtained in the classification of land use for the basin between 1986 and 2020. The results of the accuracy indices generated in the validation step are shown in
Table 4,
Table 5,
Table A1 and
Table A2.
The overall accuracy of the classifications based on Google Earth and Landsat data varied between 97.97 and 98.73%, while the kappa index values were between 0.96 and 0.97 (
Table 4), classified as excellent according to the evaluation proposed by [
49]. The general accuracy obtained for 2020, considering only the data collected in the field, was 89.44%, and the kappa index was 0.85 (
Table 5), still considered excellent for being in the range of 0.80 and 1 [
49]. The agreement between the mapping generated in this study and that of the MapBiomas project also had a satisfactory result, with overall accuracy between 91.60 and 96.56%, and a kappa index between 0.83 and 0.94 (
Table A1 and
Table A2).
In general, producer accuracy was superior for the classifications for natural areas, such as water and forest as well as for burned areas, indicating that they are more likely to be correctly classified. The user accuracy was higher for the classes for water, the Cerrado biome, burned areas, and other areas, indicating higher probability of the classified areas actually representing these categories in the field.
Table 6 presents the areas for land use and occupation obtained from these classifications, in addition to those for crops and pasture. Changes over time for these classifications are shown in
Figure 4 (1986, 1991, 2005, and 2020) and
Figure A2 (1996, 2000, 2011, and 2015), respectively.
Our land use classifications when compared to the ground truth data we collected had a relatively lower accuracy (289/394 = 73.35%) for crops compared to other land use categories, which all had >93% accuracy (
Table 5). The times that we misclassified crops as something else was highest for the pasture (75), other area (24), Cerrado (4), and forest (2), land use classes (
Table 5). Our land use classifications using Landsat data in 2020 for crops was (3280/3463 = 94.72%) (
Table A2). Misclassification as something other than crops were for other areas (143), and pasture (38). Misclassification of crops with some other land use ranged from 89.39% in 1986 (
Table A1) to 95.14% in 2005 (
Table A2).
Our results indicated growth of areas occupied by agricultural areas (i.e., pasture and crops), which led to the reduction of native vegetation (i.e., forest and Cerrado savannah) with recent stabilization to ~60% native vegetation and ~40% crops for the Teles Pires River basin (
Figure 5a).
Figure 5b presents the changes between the time periods evaluated. The percent growth in agricultural areas increased drastically by 80% from 1991 to 1996, and then by 40% from 2000 to 2005, while more recently declining by −0.6% from 2015 to 2020. This recent minor decline in total agricultural area is driven by the −10.55% drop in pasture, which takes up more land area than crops which actually increased from 2015 to 2020 (
Table 6). The percentage reduction in native vegetation has been less drastic, with greater declines from 1991 to 1996 (−9.6%) and from 2000 to 2005 (−13.3%), with stabilization to around zero with −0.2% growth from 2015 to 2020 (
Figure 5b). This has predominantly been driven by reductions in forest, since the Cerrado increased slightly by 3.05% from 2015 to 2020 (
Table 6). Between 2015 and 2020, both declines in the increase in agricultural areas and the decrease in native vegetation to near 0% (
Figure 5b) does not mean that there was no expansion of agriculture or deforestation. Rather this indicates a relative stabilization of the recent rates of agricultural expansion and deforestation.
The regionalization of land use and occupation show differences in the occupation patterns of the different parts of the Teles Pires River basin (
Figure 6). Forest was the predominant class in the basin in both years, but showed great variation of the area in the period, from 114,400 to 78,900 km
2. More than 80% of the entire mapped forest was located in the regions of the lower and middle Teles Pires, which in 1986 had forest areas corresponding to 49,200 and 49,300 km
2, respectively. The lower Teles Pires had the smallest decline in forest areas, still showing about 43,300 km
2 in 2020. In the upper Teles Pires, forests share the space with areas of the Cerrado, which also corresponds to the indigenous environment of this sub-region.
In the upper Teles Pires, forest areas showed variation between 15,700 and 7700 square kilometers (km2) of occupied area during the period, while the Cerrado areas varied between 13,000 and 5300 km2, the latter represented the class with the greatest percentage reduction in our study. Considering the forest and Cerrado classes together, in the upper Teles Pires, there was an area variation from 28,800 to 13,000 km2 between 1986 and 2020, corresponding to 84% and 38% of the region, respectively. This was the part of the Teles Pires River basin with the lowest percentage of natural vegetation in 2020.
Crops was the land use class with the greatest growth in the period. In 1986, it occupied about 2% of the basin and increased to about 16% in 2020, from 3200 to 24,100 km
2 of the area. Most of the area mapped as agriculture is concentrated in the upper Teles Pires, a region that in 2020 was more than 70% agriculture. Crops already occupy approximately 50% of the upper Teles Pires. In the middle Teles Pires, crops have increased, occupying about 10% of the region in 2020, whereas in the lower Teles Pires crops are less than 0.5% of the land area (
Figure 6). Pasture represents the predominant form of agricultural land use in the basin, covering more than 8000 km
2 in 1986, corresponding to 6% of the total area of the basin. Pastures occupied 31,000 km
2 in 2020, 22% of the land area, decreasing from the peak values of 32,385.7 and 34,643.2 km
2 that it occupied between 2005 and 2015 (
Table 6). It is in the middle Teles Pires that most of the pastures mapped are concentrated, with more than 20,000 km
2 of pasture areas in 2020, equivalent to 37% of the land area in this region.
Water also increased between 2011 and 2020, from 737 to 1159 square kilometers, a 57% growth in surface area over this 9-year period (
Figure 5c). The percentage increase from one time period to the next was more variable for other area and burned area (
Figure 5d). Formed by the junction between areas of minority territorial occupation, the classification of other area also showed an increase, primarily from the growth of urban areas in 23 municipalities within the basin. Mining areas were mostly located on the banks of large tributaries of the Teles Pires River, especially in the middle portion of the basin. Burned areas could not be analyzed. Mapping burned areas requires consideration of the temporal distribution of wildfires, using images with dates that correspond to the end of the wildfire season taking place during the dry season, with 80 to 90% of fires occurring between the months of June and October [
50].
The area conversions that occurred between land use classes were detected using a method for comparing different maps, based on the classifications of 1986 and 2020, and then the area conversion matrix presented in
Table 7 was generated. This made it possible to identify the losses experienced by each of the classes, as well as the allocation classes of these areas. In the matrix, the values in the diagonal cells indicate the area that remained in the same class of land use and occupation between 1986 and 2020, while the other values indicate the changes that occurred in the area. The rows in
Table 7 identify the losses for each land use class, while the columns correspond to the gains and their origin.
The conversion matrix makes it evident that the greatest losses of Cerrado and forest areas resulted from their conversion to crops and pasture. The forest had the highest loss of 37,900 km
2, where 23,200 km
2 was converted to pasture, and 12,700 km
2 transitioned to crops. Forest losses were also recorded due to the conversion to water. The Cerrado had the highest proportion of losses compared to the pre-existing area, with 69% being converted to other uses, representing 9000 km
2 of the lost area, of which 5800 km
2 was converted to crops and 2100 km
2 was opened to pasture. Losses of pasture area were also recorded during this period (1986 to 2020), and these were mainly due to the conversion to crops. As the main area conversions identified in the basin were from forest to crops (F/CP), forest to pasture (F/P), Cerrado to crops (C/CP), Cerrado to pasture (C/P), and pasture to crops (P/CP), these conversions were compared for all the yearly intervals mapped (
Figure 7).
Forest areas showed greater conversion to pasture for all the evaluated time periods, with higher values recorded between 2000 and 2005, when more than 10,000 km2 was converted. This period was also responsible for the greatest conversion of forest into crops. The replacement of the Cerrado areas with pasture and crops varied between the periods. From 1986 to 1991 and from 2000 to 2015, the conversion of Cerrado to pasture was higher, whereas from 1991 to 2000 the conversion of Cerrado to crops was higher, with the peak conversion recorded between 1991 and 1996. The dynamics of pasture conversion to crops was characterized by progressive growth, representing the most significant type of conversion in more recent periods.
Figure 8 shows the values of the conversion of pasture to crops over the years, mapped for all three sub-regions of the Teles Pires River basin. In the upper Teles Pires, this form of conversion from pasture to crops increased, peaking between 2000 and 2005, when about 53% of the pastures in the region were converted to crops. This type of land use conversion subsequently decreased. In the middle Teles Pires, the conversion of pastures to crops has increased, with a conversion of 3900 km
2 recorded between 2015 and 2020, representing 16% of the pre-existing pastures in 2015. In the lower Teles Pires, this type of conversion is still a recent phenomenon.