A New Climatology of Vegetation and Land Cover Information for South America

Figure 3 exhibits the new land cover maps based on the 2010–2020 climatology of the ESACCI product, reclassified for the IGBP classifications (Figure 3A) and UMD classifications (Figure 3B), respectively. Additionally, validation control points used in constructing the error matrix for various land cover classes are highlighted, totaling more than 400 points per thematic map.

Confusion/error matrices and their producer and user accuracies for IGBP and UMD classes are displayed in Table 5 and Table 6. These matrices are derived from the new ESACCI LC data spanning 2010–2020. This analysis is the primary method for evaluating mapping quality, offering an overall accuracy rate tied to Global Accuracy. It also provides accuracy and error rate estimates for each mapped class.

In Table 5, the confusion matrix analysis reveals the most significant misclassification error for IGBP in the Cropland/Natural Vegetation class, where confusion occurs with Croplands and Savannas’ classes. This confusion is easily explained when examining the Croplands class, as they are similar classes, and the coarser resolution of 0.01° may contribute to lower mapping precision in these circumstances. As for the Savannas class, representing the Brazilian Cerrado and Caatinga, classification confusion may be attributed to agricultural boundaries, deforestation, and variations in vegetation vigor due to rainfall [48,49].

The changes in land use and land cover between 2009 and 2018 are linked to the deforestation of the Cerrado and its replacement by agricultural areas [50]. This modification in land cover influences climatology because abrupt changes in land cover result in greater variability in the affected region’s coverage type. In this sense, updating this information is essential to ensure that the initial conditions of numerical models do not become outdated.

According to Table 5, the Cropland/Natural Vegetation class shows a user accuracy of 56.0% (±10.1%) with a high commission error, indicating unreliable mapping. Similarly, the Savannas class has a user accuracy of 57.4% (±6.8%). In contrast, the Cropland class demonstrates a user accuracy of 75.0% (±6.6%), considered reasonable for large, mapped areas.

Regarding producer accuracy, the values for the Cropland/Natural Vegetation, Savannas, and Cropland classes are 43.1% (±7.3%), 92.15% (±3.3%), and 72.64% (±6.1%), respectively (Table 5). Notably, only the Cropland/Natural Vegetation class exhibits a high omission error associated with incorrectly assigning the class to the pixel. In contrast, the Savannas and Cropland classes have low omission errors.

Table 5 indicates that more homogeneous land covers, such as Evergreen Broadleaf Forest, Permanent Snow and Ice, Barren or Sparsely Vegetated, and Water classes, boast user accuracy exceeding 80%, resulting in low commission errors. For omission errors associated with producer accuracy, these classes have the following values: 96.17% (±1.66%), 45.43% (±17.93%), 55.95% (±9.66%), and 100% (0), respectively. From this perspective, the Permanent Snow and Ice, Barren, or Sparsely Vegetated classes demonstrate low reliability, while the Evergreen Broadleaf Forest and Water classes exhibit high producer accuracy.

Classes like Wooded Tundra, Mixed Tundra, and Barren Tundra, with smaller coverage areas over SA, show low reliability in samples since their coverage is less than 900 km² (Table 5). Consequently, only a few sampling points (7) were generated for these Tundra classes. In the case of the Urban and Built-up class, there is a producer accuracy of only 35.5% and a user accuracy of 90%, indicating high omission and low commission errors.

The ESACCI 2010–2020 LC climatology achieved an overall accuracy of 89.83% (±0.94%) over SA, surpassing the global accuracy of ESACCI (above 80%) [1,33]. This difference is due to the coarser resolution (0.01°) of the new LC climatology. The results highlight the new LC climatology’s high applicability and precision in defining LC classes.

Table 6 displays the error matrix for the UMD classification based on the ESACCI 2010–2020 LC climatology. The Open Shrubland class shows the highest omission error (producer accuracy), signifying misclassification and confusion in land cover. This class is often mistaken for Wooded Grassland/Shrubland due to similar characteristics, as observed in Figure 4, with a user accuracy of 75.0% (±9.93%), indicating an acceptable commission error.

Table 6 reveals that the Evergreen Needle Forest and Broadleaf Deciduous Forest classes achieve perfect producer accuracy (100.0%), correctly assigned at all control points. However, user accuracy reflects the reliability of each mapped class, which shows median values of 60.0% (±24.5%) and 66.7% (±12.6%). Notably, only the Evergreen Broadleaf Forest class surpasses 90.0% (±2.4%) accuracy for both producer and user accuracy, making it the best-mapped class. For other classes, accuracy values, both for the user and the producer, hover around or fall below 80%.

The overall accuracy for the ESACCI-LC 2010–2020 climatology for UMD was 80.07% (±2.31%), based on samples over SA (Table 6). This value is very close to the global mapping of ESACCI, which reported a global accuracy slightly above 80% [1,33]. From these results, it is evident that the new LC climatology for 2010–2020 is highly applicable to the study region, demonstrating high precision in defining land use and land cover classes.

Initially, the default LC maps of the numerical models in SA classified by IGBP [34] and UMD [35] were spatially compared to the new LC maps based on ESACCI. Figure 5 and Figure 6 depict the LC maps for the two LSM classifications.

A comparison between the default land cover (LC) maps of the numerical models in SA, classified by the International Geosphere-Biosphere Programme (IGBP) and the University of Maryland (UMD), and the new LC maps based on the European Space Agency Climate Change Initiative (ESACCI) was conducted. The LC maps for the two Land Surface Model (LSM) classifications are illustrated in Figure 5 and Figure 6.

Figure 5 displays LC maps for the IGBP LC classification, revealing notable changes between the two maps. The agriculture coverage area notably expands by approximately 1,270,661.90 km² (172.79%) compared to the initial area of 735,373 km². This expansion is prominent in Brazil’s central–southern and southeastern regions, the eastern part of northeastern Argentina, and a small area in the central region of Bolivia.

It is essential to highlight that the Agriculture/Natural Vegetation class also experiences a 122.02% expansion in its coverage area (Figure 5). Consequently, there is significant degradation in these highlighted areas, with a considerable expansion of agricultural areas leading to a reduction in natural areas across SA.

Figure 5 also shows the Bare Soil Vegetation class expansions over southwestern Bolivia. However, this class decreases in size in western Argentina and a small portion of the central region of Chile. The variation in the area of this class results in a reduction of approximately 9.53% (46,833.34 km²) compared to the initial area of 491,681 km². It is worth noting that the region of the Salt Desert in Bolivia, initially classified as a water body, has been corrected to the Bare Soil class in the new LC. Other classes with significant coverage area variations include flooded areas, experiencing a 516.84% increase (150,110.51 km²) compared to the initial area of 29,044 km². Closed Shrubland also sees growth of 218.0% (260,790.03 km²) compared to the initial area of 119,630 km² (Figure 5).

The new classification emphasizes flooded areas in the Brazilian Pantanal region, a known floodable area [51,52], and in the regions of the Amazon River and its tributaries, covering Brazil and Peru (Figure 5). These changes will directly impact surface flows, influencing the pattern of variation and affecting numerical simulations over the region.

In Figure 5, the Evergreen Needleleaf Forest and Wooded Tundra classes saw an over 85% reduction from their initial coverage areas—123,886 km² and 4595 km², respectively. Conversely, the Mixed Forest, Wooded Savanna, and Permanent Snow and Ice classes exhibited approximately 75% variation compared to their initial areas—166,472 km², 671,904 km², and 24,833 km²—with only the Wooded Savanna class experiencing a reduction of about 515,000 km².

The Open Shrubland class decreased by approximately 58% (1,207,393.43 km²). Simultaneously, the Deciduous Broadleaf Forest and Mixed Tundra classes saw reductions of 18.28% (85,342.65 km²) and 20.14% (37.29 km²), respectively. It is important to note that this LC update mapped the Evergreen Needleleaf Forest and Barren Tundra classes over SA. However, their coverage areas are less than 800 km², making it challenging to identify on the map in Figure 5.

It is worth mentioning that the coverage classes, such as Evergreen Broadleaf Forest, Savannas, Pasture, and Urban and Built-up classes, had variations in coverage areas that were less than 6% of their initial areas. These areas were 7,623,213 km², 2,828,317 km², 1,594,027 km², and 44,349 km², respectively (Figure 5).

Figure 6 presents LC maps for the land cover UMD classification. Substantial differences between the maps are noted, with discrepancies in the classification of cover classes. It is relevant to highlight that the Urban and Built-up land cover class was not classified in the initial LC map, but was exclusively represented in the new LC product based on ESACCI. This class covers an area of approximately 52,397.42 km².

Additionally, there is an apparent suppression of around 2,000,000 km² in the Woodlands coverage area. Replacements for this region includes cover classes such as Deciduous Broadleaf Forest, Croplands, and Wooded Grassland/Shrublands. The class showing the greatest expansion in area is Croplands, experiencing a growth of 128.0% (equivalent to 1,309,588.84 km²) compared to its original area of 1,022,691 km². This expansion is observed in Brazil’s central–western and Argentina’s northeastern regions (Figure 6).

It is noteworthy to highlight that the expansion of Croplands primarily occurred over areas previously occupied by Wooded Grassland/Shrublands. In Brazil, this refers to the Cerrado biome, both the typical and sparse varieties [53,54]. In northeastern Argentina, this expansion is associated with the cover classes of Woodlands and Wooded Grassland/Shrublands [54,55].

The land cover classes that experienced the most significant suppression of their initial areas were Mixed Forest (−77.88%, equivalent to 68,026 km²) and Woodlands (−75.50%, equal to 2,034,272.4 km²). These classes mainly lost ground to the Croplands, Grassland, and Wooded Grassland/Shrublands cover classes [53,55,56].

There is new climatological information for GVF and LAI in SA based on GLASS products [29,30,41]. Monthly maps of the updated climatology for these vegetation variables (GVF and LAI) from 2010 to 2020 over SA are presented in Figure 7 and Figure 8. Figure 9 shows the monthly rainfall accumulations from the climatology of the MERGE product (2000–2022). In these figures, the seasonality of vegetation over SA and its response to changes in rainfall accumulations become evident. This pattern was expected as vegetation responds to water availability; when there is sufficient precipitation, vegetation exhibits greater vigor, resulting in improved plant health and increased leaf area [49,57].

In Figure 7 and Figure 8, the northeast and central–west regions of Brazil, Paraguay, southeastern Bolivia, and Argentina exhibit well-defined seasonality, with variation in the LAI and GVF values in response to periods of higher and lower rainfall accumulations, as illustrated in Figure 9. This pattern of vegetation variability was expected for these regions, as they are composed of biomes that respond rapidly to local rainfall. Among them, the Brazilian Cerrado and Caatinga stand out [2,49,57,58], and the Dry Chaco is present in Bolivia, Paraguay, and Argentina [23].

The coastal region of Brazil exhibits distinct GVF and LAI values compared to the more inland areas of the continent (Figure 7 and Figure 8). This result is attributed to higher rainfall accumulations in this region (Figure 9), as moist winds from the ocean influence it throughout the year. These winds significantly contribute to rainfall in the region [59,60,61]. The Amazon biome region, which covers Brazil, represents more than 60% of this Forest that encompasses Peru, Colombia, and Venezuela [62] and shows low variability in the GVF and LAI indices (Figure 7 and Figure 8). In general, throughout the year, GVF values consistently exceed 0.80 (Figure 7), while LAI varies between 3.5 m²/m² and values greater than 4.5 m²/m² (Figure 9).

This region of the Amazon biome is characterized by high rainfall accumulations throughout the year, leading to the absence of a dry season [62,63] and maintaining rainfall occurrence throughout all months.

These high rainfall accumulations contribute to the low variability in vegetation indices in the Amazon rainforest. Additionally, the Forest plays a crucial role in maintaining the balance between climate and vegetation. The high evapotranspiration in the forest helps sustain elevated rainfall accumulations throughout the region [62,64,65], thus supporting the local hydrological cycle.

Concerning the desert regions of SA, such as the Atacama Desert in Chile and the nearby arid areas, very low values of GVF and LAI are noticeable, with little variability throughout the months (Figure 7 and Figure 8). For GVF, values are below 0.5 throughout the year, and for LAI, index values range between 0 and 1 m²/m². Due to the arid nature of the region, rainfall accumulations are low throughout the year, not exceeding 45.5 mm/month (Figure 9).

In Figure 10, GVF scatter plots for land cover classes Cropland, Forest, Grass, and Savanna are presented, along with statistical indicators assessing input information from the MODIS numerical models for SA when compared to the new climatology based on GLASS (2010–2020).

It is observed in Figure 10 that the agreement ‘d’ of the data is low for the Forest class, with a value below 0.5. For the Cropland and Grass classes, the ‘d’ index is moderate, ranging between 0.666 and 0.718, respectively. The ‘d’ index value for the Savanna class is high, exceeding 0.927, closer to unity, and better than the other highlighted classes.

Regarding the confidence index (c), a combination of ‘r’ and ‘d’ indices, it is noted that the Forest class shows a ‘Poor’ performance, with a ‘c’ value below 0 (Figure 10). The Cropland and Grass classes have a ‘Good’ performance, while the Savanna class has an ‘Excellent’ performance. In this perspective, the Forest class requires updating of its representation in the numerical models for SA. In contrast, the Savanna class continues to be a good representation of the biome conditions. The Grass and Cropland classes still perform well, albeit lower than the Savanna class.

In Figure 11, LAI scatter plots for land cover classes Cropland, Forest, Grass, and Savanna are presented, along with statistical indicators to assess input information from the MODIS numerical models for SA when compared to the new climatology based on GLASS (2010–2020). It can be observed in Figure 11 that all land cover classes show a moderate correlation, ranging between 0.431 and 0.661. The R² is low for all classes, with a value below 0.45. Additionally, bias and MAE values for all land cover classes are relatively high, representing an average of approximately 18.3% compared to the data medium. The highest error percentile is observed for the Cropland class (36.4%), followed by Savanna (21.0%), Grass (10.7%), and Forest (5.6%).

Figure 11 shows that the agreement index (d) for Cropland and Forest classes is low, with values below 0.5. The d index is moderate for the Grass and Savanna classes, ranging between 0.641 and 0.688, respectively. Regarding the c index, all classes show ‘Poor’ or ‘Very Poor’ performance, with c values below 0.45. In this perspective, the need to update all LAI classes becomes evident, since the pattern of data variability is different. Thus, the default representation of MODIS in the numerical models of SA does not represent the region.