From the DEM, six geomorphometric attributes were generated for the study area: hypsometry, slope, curvature, flow direction, flow accumulation and topographic wetness index. For hypsometry, the values varied from 0 to 5 meters above sea level, and the ranges from 10 to 15 and > 15 meters were limited to the areas of mobile dunes. Regarding slope, 63.5% of the area has flat relief (0 to 3%). Analysis of the curvature shows that there is a predominance of rectilinear-planar segments, which correspond to 94.74% of the area. The convex-divergent and concave-convergent areas together correspond to only 5.26%, occurring in the ranges of mobile dunes.
Flow direction shows a predominance of the North (22.26%) and West (19.47%) classes, justified by the fact that the area is located in the low course of the Parnaíba River, mouth with the Atlantic Ocean, general base level. Regarding flow accumulation, sites with high-value cells indicate areas with high viability for the occurrence of drainage. In the Parnaíba River Delta, the highest values are found in areas of occurrence of hydromorphic soils and with fluvic character, to the southeast, formed under strong influence of alluvial sediments.
For the topographic wetness index, high values were observed for more saturated areas (lowland areas that accompany the river plains or areas that favor water accumulation in the soil) and lower values were observed for well-drained areas and areas with slope greater than 8% (mobile dune fields), with a variation from − 1.5 to 13.9. It is worth pointing out that the equation for obtaining TWI is Ln (a / tan (B)), where a is the ‘specific’ catchment area (that is, the upstream inflow area normalized for a measure of contour length/flow accumulation) and B is the slope gradient, in radians, in the grid cell (slope). TWI will display negative values when the catchment area is lower than tan(B), since the natural logarithm of any value lower than 1 will be negative. For the Parnaíba River Delta, flow accumulation and slope had zero values for most of the area, justifying the variation of the values obtained for TWI. Figure 4 shows the frequency histogram of these attributes.
From the images of Landsat 8’s OLI sensor, it was possible to generate the following predictive variables: band 2 – blue, band 4 – red, band 6 – medium infrared, band 10 – thermal band, CLAY, IRON, NDVI and NDWI. Band 2 showed high values of reflectance in the range of mobile dunes and in the areas of sandspit and beach, assisting in the identification of sandy soils and soils with hydromorphic character, represented by the low values of digital numbers.
Band 4 proved useful for discriminating the vegetation, as it is the red band of absorption by chlorophyll. Like band 2, band 4 showed high values of reflectance for areas of sandy soils. In sites where the vegetation is dense, such as mangrove forests and areas of swamp vegetation, there was greater absorbance in the spectral range of red. Band 6 (shortwave infrared), analyzed individually, also showed satisfactory results for a differentiation of the dune fields with the sandspit and beach areas, where the values were high. Through the interpretation of the low values it was possible to delimit the mangrove areas. Although the soils found in these sites are similar in character, these differentiations were useful in the training of the prediction model used.
The thermal band of the TIRS sensor makes it possible to remotely estimate soil temperature by transforming the digital number of the image into radiance. For the study object, pixel values ranged from 22.9 to 37.5 and were effective in the delimitation of the areas of Fluvial-Marine Plain (values between 24.1 and 25.4), being possible to find associations of hydromorphic soils (Thionic Gleysol + Umbric Gleysol + Gleyic Solonchack + Thionic Histosol), all of clayey or indiscriminate texture. The image also assisted in the identification of areas of fluvial-alluvial deposits (values between 31.2 and 37.5), representing associations mainly of soils originated from alluvial sediments with indiscriminate texture (Abruptic Solonetz + Fluvic Cambisol + Eutric Fluvisol + Sodic Vertisol).
CLAY showed digital values ranging from 0.43 to 3.0. Higher values were associated with the presence of clay minerals present in soils of areas of fluvial-marine plain and fluvial terrace, which can be correlated with hydromorphic soils. The lowest values represent soils with the presence of primary minerals, mainly sands constituted by quartz found in the areas of mobile dunes, sandspit and beach.
IRON ranged from − 22.2 to 34.1, and it was possible to identify higher values in areas of mobile dunes and fluvial-alluvial deposits, with associations of soil with sandy texture and alluvial soils with indiscriminate texture, which contain considerable amounts of primary minerals. Lower values indicate deposit areas of swamps and mangroves, suggesting the occurrence of hydromorphic soils. Similar values were verified for the areas of stabilized dune and paleodunes. In these places there is a predominance of herbaceous and/or shrubby plant species and it is common to find water bodies due to the low position in the relief and proximity of the groundwater to the surface, where Arenosols can occur and with strong restriction to drainage, that is, the presence of hydromorphism. This situation of relief causes reduction of iron and manganese oxides to their most soluble ionic forms.
For NDVI, the digital values of pixels between − 1 and − 0.10 indicated water bodies. The pixels with values from − 0.09 to 0.15 indicate areas with low vegetation response. These areas correspond to Mobile Dunes, Sandspit and Beach, consisting predominantly of poorly consolidated sandy materials. Values between 0.15 and 0.30 represent areas with presence of herbaceous and/or shrubby undergrowth and, associated with it, mainly Eutric Aerosol. The values between 0.3 and 0.4 classified the areas of open natural fields with the presence of Eutric Fluvisol. Values from 0.4 to 0.6 indicate areas with predominance of more shrub/tree sized vegetation, where it is possible to find associations of Abrupt Solonetz + Fluvic Cambisol. Values from 0.6 to 1.0 indicate the areas of fluvial-marine plains, with dense mangrove tree vegetation and swamp vegetation, as well as occurrence of hydromorphic soils.
NDWI is highly correlated with the water content in the vegetation cover and makes it possible to monitor changes in biomass and evaluate the water stress of vegetation. For the study area, the variation was from − 0.08 to 0.87, indicating the presence of areas with large water accumulation, and it was possible to find in these sites hydromorphic soils, both Gleysols in areas of fluvial-marine plain (range between 0.06 and 0.18) and Fluvisols in areas of fluvial terrace (range between 0.18 and 0.26). The values between 0.26 and 0.34 indicated areas of fluvial-alluvial deposits with vegetation of the lowland tropical forest type with the presence of Carnauba palms (Copernicia prunifera), where Abruptic Solonetz and Fluvic Cambisol occur. Between 0.34 and 0.45, it was possible to classify the wind plain, indicating the occurrence of sandy soils (Eutric Arenosol and Carbic Podzol). It is worth pointing out that the orbital image was obtained in October, coinciding with the dry period (low rainfall). In the first quarter of the year, water concentration is common in these places. The areas of sandspit, beach and mobile dunes are identified by the values between 0.45 and 0.87. Figure 5 shows the frequency histogram of the spectral attributes.
By crossing the predictive variables with soil classes, mapping units and landscape features, it was possible to form 2 data matrices, previously described. After creating the data matrices, the analyses were performed by the Weka program, and the cross-validation option with 10 partitions was selected to test the J48 algorithm through the accuracy assessment. The result of the hit rates of the classifier algorithms for each data matrix is presented in Table 3. Processing generates an automatic kappa value for each tested set.
Table 3
– Results of data matrices processing for MDS
Matrices
|
CRI (%)
|
Kappa
|
1) Map Units (CABRAL, 2018)
|
56.1
|
0.45
|
2) Landscape Features
|
75.5
|
0.66
|
Legend: CRI - Correctly Ranked Instances |
For matrix 1, the classifier model selected the parameters NDVI, IRON, thermal band, band 2, hypsometry and flow accumulation as the ones with greatest predominance for the classification. Based on the confusion matrix, it is observed that of the seventeen cells previously indicated as GLti, fourteen were correctly classified, with a hit percentage of 82.3%. Only half of the cells indicated as GLeu were correctly classified, and of the total, 37.5% were classified as SNap. A justification is that the SCgl present are located in environments similar to those of FLeu, representative of the SNap class. The SNap class show hits in seventeen cells out of a total of thirty (57%), with 20% of the remaining cells classified as Gleyc soil associations (Fig. 6).
Based on the classifier model of matrix 2 (Fig. 6), it is verified that the parameters used by the algorithm were: NDVI, IRON, thermal band, band 2, CLAY, band 4, hypsometry and curvature. NDVI was decisive for the construction of the model, being able to individually separate two classes, SB and MD, with the help of IRON. Another attribute of great representativeness was the thermal band, due to its ability to determine soil moisture, which identified the GL class.
For the determination of the other classes, the other parameters were considered in the classification. However, it is worth mentioning that the geomorphometric attributes had lower relevance, and only hypsometry and curvature were entered in the model. This information, combined with the result obtained from the analysis of the matrix 1 model, indicates that the attributes obtained from the DEM, for the study area, have low effectiveness for classification, due to the low variability. Another limitation refers to the spatial resolution of the DEM employed, so a product with higher level of detailing is necessary to verify greater correlation of morphometric attributes with soils.
The digital soil maps of matrices 1 and 2 were generated and validated by means of an error matrix that confronts the classes predicted by the model with the actual classes identified in the field at the 21 points pre-selected for this validation. The maps were also compared with the pre-existing conventional soil map, according to the procedures described above. Figure 7 illustrates the digital soil map of the Parnaíba River Delta generated through the data matrix 1.
For this processing, the largest class was AReu, which comprises an area of 89.28 km² (38.57%). In this class the Eutric Arenosols stand out, with simple grain structure. They contain the mineral quartz as predominant in all of their fractions. This class also comprise Carbic Podzol, sandy soils that have spodic horizon, with illuvial accumulation of humified organic matter, combined with aluminum, with very dark gray color (10YR 3/1) in the subsurface diagnostic horizon and black color (7.5YR 2.5/1).
The second largest class was SNap (42.11 km², 18.19%). Among the soils, the occurrence of Abruptic Solonetz stands out; these are soils with high sodium saturation, with prismatic or columnar structure. Its high textural gradient causes great susceptibility to erosion, also favored by the low permeability of B horizon. There is also the occurrence of Eutric Fluvisols, which are poorly evolved mineral soils formed from recent alluvial deposits. In addition to these, there are also Fluvic Cambisols, which consist of mineral material with cambic horizon, underlying the A horizon, with not very advanced degree of development and irregular variations of granulometry in subsurface. The last component of the association is the Sodic Vertisol, which consists of mineral material, with vertic characteristics in the horizons, appearance of cracks in the dry period due to clay shrinkage and swelling, and slickensides.
The grouped classes GLeu and GLti cover an area of 59.46 km² (corresponding to 25.68%). In this unit, the soils found have gleyic properties. These soils often have mottles or variegated colors and can assume any hues and values as long as the chroma is less than or equal to 2.
Classes MD and SB extended for 27.59 km² (11.91%) and 37.99 km³ (16.41%), respectively. Mobile dunes are large, individual moving masses of sand, consisting of simple and/or composite wind dunes, as well as large strips of sand stretched near the beach line. The sandspits and beach are accumulations of sand located between the base of the modal waves and the boundary of the beach, deposited mainly by the waves, but are also influenced by the tides and the local topography.
Regarding the accuracy of the mapping, the validation by means of an error matrix, expressed in Table 4 showed overall accuracy of 61.90%, higher than that of the previous map. The same is true for kappa, which for this map reached 0.52, a good agreement according to criteria of Landis and Koch (1977).
It is worth mentioning that, according to Ten Caten (2011), the mean value of kappa index in studies conducted in Brazil is 0.47. This value is higher than that reported in the international literature, for example, 0.39 for flat areas, reported by Hengl and Rossiter (2003).
The map generated from data matrix 2 is represented in Fig. 8. Based on the image, it was verified that the most comprehensive class for the study area was sand texture soils - AR, with about 90.31 km² (39%). The second most comprehensive class was that of soils originated from alluvial sediments, of indiscriminate texture - FL, with 60.24 km², corresponding to 26% of the area.
The class of soils with gleyc characteristics – GL represents 19.4% of the area (approximately 44.91 km²). Most of these soils are representative of the fluvial-marine plain areas (mangrove forest areas) that are subjected to strong influence from the Parnaíba River, its tributary, the Igaraçu River, and the Atlantic Ocean. Classes MD and SB extended for 22.99 km² (9.9%) and 13 km³ (5.6%), respectively.
The error matrix of the digital map generated from matrix 2 is found in Table 5. The overall accuracy was 67% and the kappa index was 0.57, indicating good agreement according to the adopted criteria for evaluating the accuracy.