MAPPING VEGETATION ON FERRUGINOUS SUBSTRATES USING ASTER AND GAMMA-SPECTROMETRY IMAGES IN THE IRON QUADRANGLE, MINAS GERAIS

The Iron Quadrangle (IQ) region in Minas Gerais is remarkably geobiodiverse, despite a long history of anthropogenic pressures such as mining and urbanization, but still lacks detailed studies on the distribution of its remaining native vegetation in diff erent substrates. In this study, we utilized Advanced Spaceborne Thermal Emission and Refl ection Radiometer (ASTER) images, besides Gamma-spectrometry (Gamma) survey data associated with existing geological mapping (GM) and extensive fi eldwork, to discriminate and quantify remnants of vegetation on ferruginous substrates in the IQ. The Maximum Likelihood (ML) algorithm was used to classify the vegetation types, thus named: open Rupestrian Field, shrubby Rupestrian Field, Capão Forest, Cerrado stricto sensu, Cerrado Field, Seasonal Forests, Pastures and Reforestation (the latter three regardless of substrate type) associated with the predominant substrates (ferruginous ironstone, phyllites, and quartzite). The use of ASTER images alone did not allow a reliable separation of ferruginous and non-ferruginous substrates, but the integration of all diff erent data (ASTER-ML + Gamma + GM) allowed the provisional mapping of the vegetation associated with ferruginous substrates, potentially ferruginous and non-ferruginous substrates. The resulting map shows that the vegetation on ferruginous and potentially ferruginous substrates cover 8.7% and 6.9% of the IQ, respectively. The detailed analysis of the distribution and fragmentation of phytophysiognomies on ferruginous substrates is of great importance for developing strategies to conserve the geobiodiversity of the IQ, and need to be further refi ned by checking and fi eld mapping by novel approaches.


1.INTRODUCTION
The Iron Quadrangle (IQ) region has a remarkable geobiodiversity (Fernandes, 2016) but lacks detailed mapping of its vegetation formations. Such mapping could provide technical support for decision-making regarding the conservation of remaining forest and Ferruginous Rupestrian Grassland formations. The region suff ers from strong anthropogenic pressures on native ecosystems, mainly as a result of mining and urbanization (Jacobi et al., 2007).
In the mountainous regions, Lateritic Rupestrian Fields, also termed Ferruginous Rupestrian Fields (or Grasslands), or "canga vegetation," develop on ferruginous substrates (Vincent, 2004;Viana and Lombardi, 2007). These unique ecosystems, colonized by specialist plants adapted to oligotrophic environments, are capable of tolerating a number of severe environmental fi lters, such as shallow soils, severe water defi cits, low fertility, high oxidized iron concentrations, and low water retention, as well as large daily thermal amplitudes, frequent fi res, high sun exposure, and constant winds (Vincent, 2004;Jacobi et al., 2007;Schaefer et al., 2016). The vegetation shows several anatomical, morphological, physiological, and reproductive adaptations that allow it to survive in these environments (Alves and Kolbek, 1994).
The plant communities in the IQ vary as a function of the substrate type, so that the spectral attributes of soils and surface rocks, in some cases, may assist in mapping the vegetation. Remotesensing techniques, by identifying subtle changes in vegetative cover, can enable the identifi cation of changes in substrate conditions, establishing the rock-soil-vegetation association (Almeida Filho, 1984). The advent of Advanced Spaceborne Thermal Emission and Refl ection Radiometer (ASTER) sensor images has enabled discrimination of geological as well as vegetation targets (Lima et al., 2005;Gil et al., 2014). Some studies (Rouskov et al., 2005;Rajendran et al., 2011) used satellite sensors (Landsat Thematic Mapper and ASTER-Terra) for the identifi cation and discrimination of iron-rich deposits through the composition of multispectral indexes. However, this technique becomes limited when such regions are vegetated, which minimizes the eff ects of the energy refl ected by the ferruginous substrates.
Additionally, the gamma-spectrometric (Gamma) data obtained by aerogeophysical surveys enables the elimination of the eff ects of vegetation cover and the direct discrimination of ferruginous substrates by inferences regarding the geochemical characteristics of the rocks. Also, Gamma spectrometry responds to concentrations of potassium (K) radioisotopes, uranium (U) and thorium (Th) series radioisotopes. In rocks and soils; these concentrations are directly proportional to the intensity of the gamma radiation emitted by their radioactive decay (Wilford et al., 1997;Santos et al., 2008), and are frequently associated with the geochemical signature of substrates. This allows, for example, the separation of ferruginous and nonferruginous substrates, with diff erent geochemical compositions. However, the use of Gamma data has some limitations in terms of distinguishing certain substrates based on their similar responses, unresolved radioactive barriers, or diff erences in soil moisture; therefore, these data should be used with caution, and preferably combined with all available information (Wilford et al., 1997) in regolith studies.
In this sense, this work aimed to apply remotesensing techniques associated with geological data (Lobato et al., 2005) with ASTER images and Gamma data to discriminate vegetation remnants on ferruginous substrates and other substrates, in the IQ region.

2.1.Study area
The IQ is located in central Minas Gerais, with 7,800 km 2 in area. All procedures were conducted considering a 5-km buff er from the IQ limits ( Figure 1). According to a geological cartography by Lobato et al. (2005), approximately 6.4% of the IQ consists of iron-rich formations (ferruginous substrates) that are associated with generally shallow soils, where a predominant rupestrian vegetation cover ranges from fi eld to cerrado to upper montane forest. A range of characteristic soils can be found in these areas, showing great landscape and geoenvironmental diversity with the occurrence of a ferruginous substrate (Schaefer et al., 2008, Schaefer et al., 2016. The denomination "ferruginous substrate," in this study, is broad and more comprehensive than the word "canga", and comprises a range of substrates (fresh rock, altered rock, sedimentary cover, canga, and soils) associated with the following lithotypes in the geological map (Lobato et al., 2005): canga, detritallateritic cover, , iron oxide supergene concentrations, hematite bodies, ferruginous dolomites, iron formations, hard hematite, itabirite, dolomitic itabirite, laterite and ferruginous detritus, hard hematite lenses, magnetite, high content iron ore, hematite ore, residual lateritic soil, and associated colluvium and eluvium. However, these lithotypes may occur as secondary component in non-ferruginous geological units, so that the term "ferruginous substrate" is not limited to the delineation of the aforementioned lithotypes.

2.2.Pre-processing of ASTER images
Ten ASTER scenes were used, covering the 5-km IQ buff er ( Figure 1). ASTER bands 1 to 9 were those processed comprising the visible and near-infrared (VNIR) and shortwave infrared (SWIR) wavelengths. The SWIR bands, with a 30-m spatial resolution, were resampled using the nearest neighbor method to 15 m for compatibility with the VNIR bands (15 m). The images were georeferenced from terrestrial control points and orthorectifi ed GeoEye images; the fi nal data were projected in zone 23 K SIRGAS2000 UTM. ASTER images were then converted to radiance values based on the maximum and minimum radiance values of each band. These procedures were performed using ArcGIS 10 software. Then, the Fast Line-of-sight Atmospheric Analysis of Hypercubes (FLAASH) method of the ENVI 5.0 software was used for the atmospheric correction of the ASTER images.

2.3.Supervised classifi cation
In addition to bands 1 to 9 of the ASTER images, multispectral indexes were included in the supervised classifi cation, in which the following band ratios were considered to show certain spectral patterns. 1) To show areas with ferruginous substrates, the red-greenblue (RGB) band composition R = band 2/band 1, G = band 4/band 3, and B = band 4/band 5 (Rouskov et al., 2005), was used and 2) to aid in vegetation mapping, the Normalized Diff erence Vegetation Index (NDVI) was used ((band 3 -band 2) / (band 3 + band 2)) (Rouse et al., 1974). The Maximum Likelihood algorithm (ML) was used and the following vegetation cover classes were separated and subdivided as a function of the substrate type (ferruginous, quarzitic, and phyllitic): Open Rupestrian Field (ORF), Shrubby Rupestrian Field (SRF), Capão Forest (CF), Cerrado Field (CF), and Cerrado Stricto Sensu (CSS). Training and validation samples were collected from all phytophysiognomies based on geological data (Dorr, 1969;Lobato et al., 2005) and associated with the following substrates: ferruginous, quarzitic, or phyllitic (including serpentinites, metabasalts, and shales).
A large part of the area, totaling 894 control points, was covered to collect training and validation samples using a Garmin 60CSx Global Positioning System receiver. In areas of diffi cult access, samples were collected as fi eld truths based on the interpretation of high-resolution images (GeoEye) of the Native Vegetation Coverage Map of the State of Minas Gerais (Scolforo et al., 2006) and through queries to the Google Earth image bank. Additionally, samples were collected in regions where the following vegetation classes had no association with lithology: lakes, pastures, and reforestation (Eucalyptus and/or Pinus). Training and validation samples were collected in a polygon format consisting of 1,152 and 775 polygons, respectively. Classes of urban areas and areas infl uenced by mining were not included in the supervised classifi cation and were overlapped after fi nal processing via manual mapping over GeoEye images.
To evaluate the classifi cation accuracy, validation samples were used and the Kappa index and Global Accuracy were calculated. The Kappa index is a statistic that measures the agreement between fi eld truths and a classifi ed map, noting the map's legitimacy (Congalton, 1991). These values were classifi ed as proposed by Monserud and Leemans (1992). From only the training samples, the Kappa index was evaluated for all band and index combinations to identify the infl uence of each band or index on the classifi er settlement. In addition, following the classifi cation, 40-pixel (~1-ha) groupings were eliminated and replaced by the most representative neighboring pixels. For all these operations, ArcGIS 10 software was used.

Gamma-spectrometry Survey
Continuous surfaces were generated through interpolation for data enhancement using the minimal curvature algorithm of the radioisotopes K, eTh, and eU to highlight the lateritic covering and colluvium of ferrous formations (Boyle, 1982;Wildford et al., 1997). Ternary radioisotopic images were also generated using the C(K)M(eTh)Y(eU)K color scale. The ternary Gamma image was cropped to the 5-km IQ buff er ( Figure 1) and an unsupervised classifi cation was performed using the IsoCluster method resulting in 10 classes. Then, the classes were superimposed onto the available survey geological units (Lobato et al., 2005) and correspondence with the ferruginous and nonferruginous substrates was visually assessed, allowing the image to be reclassifi ed from 10 to only 2 classes by grouping. Thus, the classes obtained during the Gamma analysis were used to complement and validate the separation of ferruginous and non-ferruginous substrates from the phytophysiognomic classes obtained from the classifi cation of each ASTER scene. These procedures were performed using the ArcGIS 10 software.

Geological Data
From the available IQ geological map (Lobato et al., 2005), two additional characteristics were processed and used for data refi nement and integration as follows : 1) (Lobato et al., 2005). 2) Excluded geological units with no Fe-rich formations: which refer to geological formations without the presence of banded iron formations, used as references to erroneously reclassify the mapped areas as ferruginous substrates. These were the Bação Complex, Belo Horizonte Complex, Bonfi m Complex, Caeté Complex, Guanhães Complex, Santa Bárbara Complex, Caraça Group (Moeda Formation, only the core area of Caraça Mountain), Gnaisse Souza Noschese Unit, Granito Borrachudos Unit, Granito Peti Unit and Rocha Intrusiva Unit (Lobato et al., 2005). This fi lter was mainly used for areas where there was no Gamma data (Figure 1), i.e. the north-northwest and north-northeast regions of the study area.

3.1.Supervised classifi cation and validation
Scene 070826 (02), which covers most of the IQ (Figure 1) and has no cloud cover, was used as a reference for the supervised classifi cation. It was the basis for proposing the best band combinations and multispectral indexes for the Kappa index evaluation. Classifi cation was performed considering all combinations of individual bands and multispectral indexes comprising the best Kappa indexes. Thus, six ASTER bands comprising the best Kappa indices -namely bands 1, 2, 3, 4, 6, and 7 -were selected. ASTER sensor bands 5, 8, and 9 were not included because they did not provide combinations with high Kappa indexes. Band ratios were also selected to show ferruginous substrates according to Rouskov et al. (2005) and the NDVI. The band 4/band 5 ratio, suggested by Rouskov et al. (2005), did not change the Kappa index value, thereby also showing that it provided no contribution to the classifi cation. Thus, this was the only ratio not adopted for the other scenes.
Thus, by evaluating the confounding matrix of the reference scene, it was observed that, in general, the Ferruginous Rupestrian Fields (Open and Shrubby) were more spectrally separated from the Rupestrian Fields compared to that of the other lithologies, that is, they were better identifi ed. In this regard, a cluster was proposed in which the phytophysiognomies on ferruginous substrates were maintained and the phytophysiognomies on non-ferruginous substrates were grouped (quartzites and phyllites), which improved the Kappa index in all scenes. According to Monserud and Leemans (1992), Kappa values obtained for the fi nal classifi cation of this grouping were considered reasonable, good, or very good except for those scenes with lower spatial expression (Table 1 and Figure 1). For the fi nal analysis of the supervised classifi cation, the Kappa index was calculated through the validation samples for the entire study area considering only the OBS: Scenes 040824 (07) and 040824 (08) have vegetation classes only on ferruginous substrates and on quartzites; thus, no grouping is necessary.

Scenes
Kappa Indexes/ Global Accuracy vegetation samples on ferruginous and non-ferruginous substrates. The fi nal Kappa index obtained was 0.61 and the overall accuracy was 0.64, showing good agreement between the fi eld truths and classifi ed map according to Monserud and Leemans (1992).

3.2.Integration of data
The integration of ML classifi cation data with Gamma and geological data allowed the separation of three substrate types: ferruginous, potentially ferruginous, and non-ferruginous. Figure 2 shows a synthesis of data integration used as a diff erentiating criterion between ferruginous and potentially ferruginous substrates. After integrating the data, areas in the geological map considered as Fe-rich lithotypes and those indicated as ferruginous substrate in the three information sources were considered areas with vegetation on a ferruginous substrate. In a complementary manner, the ferruginous substrate polygons indicated by the integration of the data extrapolating and intercepting the ferruginous lithotypes presented in the geological map (Lobato et al., 2005) had their limits considered as the vegetative cover on the ferruginous substrate.
Areas with a potentially ferruginous substrate were associated with a positive indication in two databases, i.e., the ASTER images and the Gamma classifi cations; by the ASTER image classifi cation and the geological map as a unit with iron formations at the subordinate level; or in the classifi cation of Gamma data and the geological map as a unit with subordinate iron formations. The areas with a non-ferruginous substrates correspond to the remaining areas indicated and also excluded areas of other geological units such as gneisses, granites, and other associations with no known occurrence of Fe-rich formations, particularly for areas without Gamma coverage (Figure 1).

Quantifi cation of vegetation on ferruginous substrates
The remaining areas of ecosystems developed on a ferruginous substrate cover 676.9 km 2 (8.7%) of the IQ (Table 2 and Figure 3). The most representative phytophysiognomies are the Capão Forest /Seasonal Forest and the Shrubby Rupestrian Field followed by the Open Rupestrian Field, Cerrado Stricto Sensu, and Cerrado Field, respectively ( Table 2). The areas impacted by anthropogenic activity are the urban areas, areas aff ected by mining, pastures, and reforestation which together correspond to 24.6% of the total IQ area, proving the great pressure exerted on the natural ecosystems of this region.

4.DISCUSSION
The strong anthropogenic pressure on vegetation remmants associated with the ferruginous substrates in the IQ region (Jacobi et al., 2007), which represent extremely important areas for biodiversity conservation in the transition between two large Brazilian hotspots, the cerrado and Atlantic forest (Myers et al., 2000;Fernandes, 2016), justifi es detailed studies of the distribution and conservation degree of these phytophysiognomies. In this sense, the use of remote sensing techniques enables identifi cation of these phytophysiognomies and associated substrates (pedological and geological) (Yamaguchi and Naito, 2003;Kalinowski and Oliver, 2004;Rouskov et al., 2005;Rajendran et al., 2011). For some remote orbital sensors obtained through refl ectance signals from objects on the Earth's surface (passive sensors), vegetation can interfere or hinder substrate identifi cation, particularly in a closed canopy. However, other sensors, such as gamma spectroscopy, can penetrate vegetation and obtain geochemical data from the upper 30 cm of soils (Minty, 1997;Wilford et al., 1997). Regarding the images with spectral data from the ASTER sensor, the combination of bands with the best Kappa indexes-initially calculated with the training samples only-includes six ASTER bands. ASTER sensor bands 5, 8, and 9 were not included because they were not predominant among combinations with higher Kappa indexes. Rowan and Mars (2003) suggest a good spectral separation of lithological categories for ASTER sensor bands 5 to 9 except for ferruginous deposits. The VNIR bands (Bands 1, 2, and 3) have information regarding metal absorption, particularly iron (Adams, 1974;Rowan et al., 1995), and chlorophyll absorption during vegetation photosynthesis (Knipling, 1970). These data corroborate the greater importance of these bands for mapping these phytophysiognomies. In this regard, the grouping of phytophysiognomies in classes on ferruginous substrates and diff erentiated phytophysiognomies on non-ferruginous substrates (quartzites and phyllites) is also important for the improvement in the Kappa index in all scenes (Table 1), as it also favors the aspects that show phytophysiognomies related to the ferruginous substrates.
In addition to the combination of ASTER bands, the inclusion of band ratios to highlight the ferruginous substrates (Rouskov et al., 2005) and NDVI were also chosen as those with the highest Kappa indexes. The 4/5 band ratio was also noted by Bierwirth (2002) as important in the identifi cation of laterites although it was not infl uential in this study. NDVI signifi cantly contributes to the identifi cation of diff erent vegetation types as it is infl uenced by the productivity and photosynthetic dynamics of phytophysiognomies (Rouse et al., 1974;Petorelli et al., 2005).
The defi nition of areas with greater uncertainties, defi ned as potentially ferruginous substrates, reinforces the diffi culty of mapping these vegetation typologies using spectral data, particularly when associated with specifi c substrates, as in the IQ region where there are great variations in geological units with diff erent degrees of metamorphism (Lobato et al., 2005). Thus, Gamma data, in addition to assisting geological and pedological surveys (Vasconcellos et al., 1994;Wilford et al., 1997;McBratney et al., 2003;Santos et al., 2008), can be extremely important for mapping phytophysiognomies associated with diff erent substrates. Therefore, this mapping allows for a more comprehensive analysis of the distribution and fragmentation pattern of the phytophysiognomies on ferruginous substrates which may generate information of great importance for future planning. These data can be combined with more in-depth fl oristic studies to support future IQ geobiodiversity conservation strategies and provide a more reliable predictability of the impacts caused by anthropogenic actions on these ecosystems. Additional work, such as reducing data using Principal Component Analysis or using more specifi c remote sensors such as Magnetometry, is required to improve the classifi cation accuracy, particularly for a more precise defi nition of potentially ferruginous areas. These results should be taken with caution due to confusing identifi cation of potentially ferruginous substrates, and further vegetation mapping combined with fi eld recognition, must be done, employing novel approaches.

5.CONCLUSIONS
Selected ASTER sensor bands, band ratios for ferruginous substrates, and NDVI were combined with Gamma and geological data allowed a consistent and reliable vegetation mapping of ferruginous, nonferruginous and potentially ferruginous substrates, the latter with lower confi dence level. This mapping shows the complexity of the study region and allows detailed analysis of the distribution and vegetation fragmentation pattern of ferruginous substrates, which is of great importance for conservation strategies in the remarkably geobiodiverse IQ region in Minas Gerais, one of the largest iron mining area worldwide.