Integrating Tacit Knowledge and AI for Digital Soil Mapping in Eastern Amazonia: Ensemble Learning, Model Performance, and Uncertainty Incorporation

Abstract

Predictive Digital Soil Mapping (PDSM) in Eastern Amazonia faces challenges due to its environmental complexity, difficult access, and scarce legacy data. While legacy soil maps contain valuable tacit knowledge, updating them requires methods that can handle uncertainty. This study evaluates the integration of old soil maps with machine learning to update soil information in Tracuateua, Pará, with a specific focus on the performance of ensemble learning and the explicit incorporation of uncertainty metrics in soil mapping units under hydromorphic influence, which, in addition to being difficult to access, are influenced by complex pedogenetic processes. We combined 270 sampling points, equivalent to the total pixels that captured the variability of soil mapping units, with environmental covariates and historical data. Several algorithms were tested, including an ensemble approach, to predict mapping units and quantify uncertainty through entropy and confusion indices. The ensemble model demonstrated improved stability and reduced classification uncertainty compared to single models, particularly in challenging hydromorphic environments. Although accuracy gains were modest, the models captured soil–environment relationships, with climate as: Annual Mean Temperature 22,000 years ago (Tmean_22k), relief: Channel Network Base Level (CNBL and altitude) and organism variables: Land Surface Temperature (LST) emerging as the main predictors. Spatialized uncertainty estimates, expressed through entropy and the confusion index, provide a practical decision-support tool for guiding field surveys and identifying areas of low mapping reliability. By explicitly transferring the pedologist’s mental model—encoded as tacit knowledge in legacy soil maps—into ensemble learning, this study presents a robust and transferable framework for updating soil maps in data-scarce tropical regions, balancing predictive performance, spatial consistency, and uncertainty-aware interpretation.

1. Introduction

Knowledge of the spatial distribution of soils is of fundamental importance for supporting the management of environmental resources such as the atmosphere, river basins, and soil itself [1,2]. In tropical regions such as the Amazon biome, soil surveys can provide essential information for the development of projects and the planning of soil use, management, and conservation [3]. However, maps with sufficient detail to guide decision-making in the Amazon are still scarce, and the lack of high-resolution information remains one of the main limitations for sustainable agricultural planning and the implementation of conservation policies in this biome.

In Brazil, detailed soil surveys at a scale of 1:100,000 or larger cover less than 6% of the entire territory, according to information from the National Soil Program (Pronasolos). Furthermore, in the Brazilian Amazon, which covers approximately 5 million km², the vast majority of available surveys are at a smaller, exploratory, or low-intensity reconnaissance scale (1:1,000,000 to 1:250,000). Among the challenges hindering the slow progress of soil surveys in this biome since the late 1980s, when the RADAMBRASIL project was completed, are the vast territory and poor accessibility (except via major waterways) [4], which demands time and considerable availability of human and financial resources [5].

In this context, many studies have addressed the possibility of not only increasing the level of detail in legacy maps using predictive digital soil mapping (PDSM) techniques [6], but also incorporating uncertainty into these estimates [2]. The formalization of the PDSM, based on the SCORPAN approach published by [7], is performed by adding the spatial position (n) to the soil formation factors formulated by [8]. It allows for the prediction of soil distribution patterns or attributes in the landscape based on their relationship with environmental covariates through mathematical and statistical methods, generating a spatial model [6].

Thus, PDSM consists of developing an empirical predictive model using statistical and mathematical methods, which combines, in short, environmental information and the tacit knowledge of soil scientist, the latter materialized in conventional soil survey maps [9]. This predictive model, however, is also capable of learning the implicit uncertainty of soil classes from legacy maps [10]. The uncertainty of the predictions is related, among other causes, to the quality of the database and the variability of the soil [11]. In this sense, considering pixel-by-pixel conversion from a soil mapping unit, in which the target is classified according to its similarity to the object, it can be evaluated by the degree of certainty in a pixel classification, which is associated with a specific soil class or attribute [11].

The possibility of improving mapping quality based on legacy maps has been addressed in a literature review of 90 publications related to digital land mapping and modeling [12] and in a study [13] exploring the potential of machine learning to accelerate the mapping of parental soil material in the northern region of Europe. In our work, we propose improvements in the Amazon biome. In traditional mapping, among soil formation factors, relief is the factor that strongly depends on the soil scientist’s tacit knowledge and the subjective interpretation of the morphogenetic criteria used to establish soil–landscape relationships [14,15]. Ref. [16] also highlights how protocols such as the Cornell method, along with other systematic evaluation approaches, reveal the subjectivity and variability in the quality of historical information contained in traditional soil survey maps, whose cartographic generalization results in greater spatial uncertainty [17].

When combined with field pedological knowledge and important pedogenetic predictive covariates [18], PDSM becomes a more efficient, reproducible, and economical approach to reduce soil information gaps, like in the Amazon biome, where detailed scale soil surveys are widely scarce and economically challenging in the context of the territorial dimension of the Brazilian Amazon. This approach helps to overcome the limitations associated with the subjectivity of tacit knowledge applied in traditional soil surveys [19] and has shown particular promise in addressing the challenges posed by traditional maps, whose spatial and textural accuracy often does not correspond to their nominal publication scale, thus limiting their usefulness for detailed assessments [16].

The use of machine learning-based modeling techniques—integrating field samples, pedological knowledge, and interrelated environmental variables—constitutes the conceptual core of PDSM [20]. Among the algorithms most commonly applied in pedometric studies are Random Forest, Ranger, XGBoost, and C5.0, all based on decision tree structures. These methods have become fundamental for PDSM due to their ability to model complex nonlinear relationships, handle high-dimensional predictor spaces, and capture interactions between multiple environmental covariates [21,22,23]. Furthermore, ensemble learning approaches have gained increasing prominence by combining predictions from multiple algorithms, resulting in more robust models with lower variance and superior predictive performance compared to individual learning [24,25,26].

This study addresses the improvement of a soil map published in 1998 for the municipality of Tracuateua, with a scale of 1:100,000, in northeastern Pará, by integrating PDSM, environmental covariates, and multiple machine learning algorithms, including ensemble modeling. Its innovation lies in improving the detail of soil maps for the Amazon biome, focusing on soils in hydromorphic environments, known for their high pedogenetic complexity and difficulty of access. It explicitly examines the implications of using legacy maps as training labels and the spatial dependence of the models in the context of evaluating the uncertainty of the predicted maps.

Given this, our hypothesis is that ensemble modeling increases the stability of predictions and reduces the spatial uncertainty in predicting soils under hydromorphic environments compared to individual models. Therefore, the objectives of this study were (i) to compare the predictive performance of different machine learning algorithms, (ii) to analyze the effect of ensemble modeling on the stability of predictions, and (iii) to quantify and interpret the spatial uncertainty of classification, with an emphasis on soils under critical areas for field surveys such as in hydromorphic environments.

2. Materials and Methods

2.1. Study Area

The study area encompasses the municipality of Tracuateua, covering 828.025 km² and located within the Amazon biome in the northeastern Pará mesoregion, between approximately 01°02′ S latitude and 46°56′ W longitude (Figure 1). The regional climate is classified as Am—tropical monsoon—according to the Köppen system, with a mean annual precipitation of 2581 mm and a mean annual temperature of 27.8 °C [27]. This tropical climatic regime strongly influences soil weathering processes, acting in conjunction with the parent material and the variability of geomorphological positions that structure the local landscape [3].

The study area is characterized by the predominance of sediments from two geological periods. (i) The Quaternary period is represented by recent alluvial deposits along the banks of large rivers such as the Japerica River, as well as beaches and mangroves. In these hydromorphic environments, GLEISSOLOS HÁPLICOS, NEOSSOLOS QUARTZARÊNICOS Hidromórficos, and PLINTOSSOLOS, according to the Brazilian Soil Classification System (SiBCS), represent common soil classes in Amazonian hydromorphic landscapes (floodplains, lowlands, and transition areas); and (ii) The Tertiary period is represented by the Barreiras Formation, which occupies approximately 70% of the study area and forms a lower Amazonian plateau characterized by LATOSSOLOS and ARGISSOLOS, predominate under flat to gently undulating relief. These highly weathered soils typically exhibit yellow to red hues associated with the presence of iron oxides, which exert strong control over their spectral behavior. Soil reflectance in the visible (VIS) and near-infrared (NIR) regions is primarily governed by soil color and mineralogical composition, particularly iron oxides, clay minerals, and organic matter. The dominance of hematite and goethite in these tropical soils results in characteristic spectral patterns and reflectance variations in the VIS–NIR range. In contrast, soils formed under hydromorphic conditions—where prolonged water saturation promotes iron reduction and organic matter accumulation—generally exhibit lower reflectance and weaker spectral features, often associated with grayish or bluish colors typical of gleyed environments [28,29,30].

2.2. Data Sources

The database is composed of mapping units (MUs) defined in the conventional soil map of the municipality of Tracuateua, originally produced by [31] at a scale of 1:100,000. This legacy soil map was published at the Great Group level, according to the SiBCS, and grouped into 9 MUs (

Table 1. Mapping units from soil classes derived from the legacy map levels of detail (LD1 and LD2) and their correspondence to the WRB/FAO classification.

Mapping Units from Legacy Map (SiBCS)	WRB/FAO	Mapping Units Adapted	Level of Detail		Area (km2)
			LD1	LD2
	Arenosols	NEOSSOLOS	R	RQo1	58.18
NEOSSOLOS QUARTZARÊNICOS Órticos + ARGISSOLOS VERMELHO-AMARELOS Distróficos	Arenosols + Acrisols			RQo5	107.79
NEOSSOLOS QUARTZARÊNICOS Órticos + NEOSSOLOS QUARTZARÊNICOS Hidromórficos + GLEISSOLOS HÁPLICOS Tb Distróficos	Arenosols + Gleysols			RQg5	3.27
NEOSSOLOS QUARTZARÊNICOS Hidromórficos	Fluvisols			RQg1	74.17
GLEISSOLOS HÁPLICOS Ta Eutróficos	Gleysols	GLEISSOLOS	G	GXve2	56.18
GLEISSOLOS SÁLICOS Sódicos	Solonchaks			GZn1	177.33
ESPODOSSOLOS FERRILÚVICOS Órticos	Podzols	ESPODOSSOLOS	E	ESo	45.59
LATOSSOLOS AMARELOS Distróficos	Ferralsols	LATOSSOLOS	L	LAd	212.26
ARGISSOLOS AMARELOS Distróficos	Acrisols	ARGISSOLOS	P	PAd	102.94

).

To enhance the predictive performance of the PDSM models, the original MUs were reorganized into two levels of detail:

(i) LD1, the least detailed level, comprising five mapping units composed of simple taxonomic units and associations at the Soil Order level derived from the legacy map; and (ii) LD2, the most detailed level, comprising nine mapping units composed of simple taxonomic units and associations at the Great Group level present in the original legacy soil map (Figure 2a,b).

This hierarchical structuring aimed to evaluate model sensitivity under different taxonomic resolutions and to explore the effects of class aggregation on classification accuracy and uncertainty.

For model training, 270 pixels of 30 m × 30 m each (900 m²) were selected, aiming to guarantee the spatial variability of the pedons at both detail levels LD1 and LD2, thus covering the entire study area (828 km²). Field surveys using an auger, at a depth of 0 to 1.50 m, were carried out in the area of the aforementioned 270 samples with the objective of, in addition to confirming the MUs described in the legacy soil map (

Table 2. Distribution of digital samples across LD1 and LD2.

LD1	No. of Digital Samples
R	109
G	62
E	33
L	36
P	30
Total	270
LD2
RQo1	28
RQo5	36
RQg5	23
RQg1	22
GXve2	30
Eso	33
Lad	36
PAd	30
GZn1	32
Total	270

), avoiding transition zones between the MUs (Figure 3 and Figure A1 in Appendix A), thus establishing a robust protocol to guarantee the independence between the MUs training. This method sought to overcome the challenges of sparse regional inference, which is due to the logistical challenge of soil surveying in the Amazonian environment. This approach allows for the identification of areas where the uncertainty of the prediction, such as the entropy and confusion index, may require additional field investigations.

2.3. SCORPAN Covariates

Fifteen predictor covariates were pre-selected based on their pedological relevance to the MUs in the study area (Table 3), aiming to align machine learning techniques with knowledge discovery approaches [10]. These covariates represent soil-forming factors following the SCORPAN framework [7]: S—soil properties or classes; C—climate; O—organisms; R—relief; P—parent material; and A—time (age).

The climate factor was represented by annual average precipitation (AAP) and paleoclimatic variables PR_22K and Tmean_22K, derived from data estimated 22,000 years ago [6]. Climate is a key driver of soil formation, and tropical conditions promote the development of highly weathered soils. Relevant datasets and models for the study area are available from Google Earth Engine (GEE) and WorldClim at 1–6 km resolution [32,33].

Table 3. SCORPAN covariates used for predictive modeling based on pedological relevance.

SCORPAN Factors	Abbreviation	Description	Spatial Resolution/Scale	Source	References
Climate (C)	AAP	Average Annual Precipitation	6 km	Google Earth Engine. ImageCollection (‘UCSB-CHG/CHIRPS/DAILY’). Time series from 1981 to 2023	[34]
	PR_22K	Annual Precipitation 22,000 Years Ago	1 km	https://geodata.ucdavis.edu/climate (Accessed on 31 July 2024)
	Tmean_22K	Annual Mean Temperature 22,000 Years Ago	5 km	https://geodata.ucdavis.edu/climate/cmip5/2_5m/ (Accessed on 31 July 2024)
Organism (O)	SAVI	Soil-Adjusted Vegetation	30 m	Multispectral data. LANDSAT 8—sensor OLI (U.S. Geological Survey). Equation: (NIR-RED)/(NIR + RED + L) $\times$ (1 + L)	[35]
	LST	Ground Surface Temperature—during the day	1 km	Google Earth Engine. ImageCollection (“MODIS/061/MOD11A1”). Time series from 2000 at 2024	[36]
Relief (R)	CNBL	Channel Network Base	30 m	Copernicus GLO-30 (DSM) and ANADEM (DTM). SAGA GIS 9.3.1	[37]
		Level
	DD	Drainage Density		Copernicus GLO-30 (DSM) and ANADEM (DTM)	[38]
	MRRTF	Multiresolution Index of	30 m	Copernicus GLO-30 (DSM) and ANADEM (DTM). SAGA GIS 9.3.1	[39]
		Ridge Top Flatness
	MRVBF	Multiresolution Index of	30 m	Copernicus GLO-30 (DSM) and ANADEM (DTM). SAGA GIS 9.3.1	[39]
		Valley Bottom Flatness
	PFC	Profile Curvature	30 m	Copernicus GLO-30 (DSM) and ANADEM (DTM). SAGA GIS 9.3.1	[40]
	RSP	Relative Slope Position	30 m	Copernicus GLO-30 (DSM) and ANADEM (DTM). SAGA GIS 9.3.1	[40]
	Altitude	Level	30 m	Copernicus GLO-30 (DSM) and ANADEM (DTM)	[40]
Parent material (P)	Geology		1:250,000	IBGE (2023)	[40]
	IOR	Iron Oxide Ratio	30 m	Multispectral data. LANDSAT 8—sensor OLI (U.S. Geological Survey). Equation: (Band 4-Red/Band 2-Blue)	[41]
Age (A)	GEOM	Geomorphons	30 m	Copernicus GLO-30 (DSM) and ANADEM (DTM). SAGA GIS 9.3.1	[42]

Table 3. SCORPAN covariates used for predictive modeling based on pedological relevance.

SCORPAN Factors	Abbreviation	Description	Spatial Resolution/Scale	Source	References
Climate (C)	AAP	Average Annual Precipitation	6 km	Google Earth Engine. ImageCollection (‘UCSB-CHG/CHIRPS/DAILY’). Time series from 1981 to 2023	[34]
	PR_22K	Annual Precipitation 22,000 Years Ago	1 km	https://geodata.ucdavis.edu/climate (Accessed on 31 July 2024)
	Tmean_22K	Annual Mean Temperature 22,000 Years Ago	5 km	https://geodata.ucdavis.edu/climate/cmip5/2_5m/ (Accessed on 31 July 2024)
Organism (O)	SAVI	Soil-Adjusted Vegetation	30 m	Multispectral data. LANDSAT 8—sensor OLI (U.S. Geological Survey). Equation: (NIR-RED)/(NIR + RED + L) $\times$ (1 + L)	[35]
	LST	Ground Surface Temperature—during the day	1 km	Google Earth Engine. ImageCollection (“MODIS/061/MOD11A1”). Time series from 2000 at 2024	[36]
Relief (R)	CNBL	Channel Network Base	30 m	Copernicus GLO-30 (DSM) and ANADEM (DTM). SAGA GIS 9.3.1	[37]
		Level
	DD	Drainage Density		Copernicus GLO-30 (DSM) and ANADEM (DTM)	[38]
	MRRTF	Multiresolution Index of	30 m	Copernicus GLO-30 (DSM) and ANADEM (DTM). SAGA GIS 9.3.1	[39]
		Ridge Top Flatness
	MRVBF	Multiresolution Index of	30 m	Copernicus GLO-30 (DSM) and ANADEM (DTM). SAGA GIS 9.3.1	[39]
		Valley Bottom Flatness
	PFC	Profile Curvature	30 m	Copernicus GLO-30 (DSM) and ANADEM (DTM). SAGA GIS 9.3.1	[40]
	RSP	Relative Slope Position	30 m	Copernicus GLO-30 (DSM) and ANADEM (DTM). SAGA GIS 9.3.1	[40]
	Altitude	Level	30 m	Copernicus GLO-30 (DSM) and ANADEM (DTM)	[40]
Parent material (P)	Geology		1:250,000	IBGE (2023)	[40]
	IOR	Iron Oxide Ratio	30 m	Multispectral data. LANDSAT 8—sensor OLI (U.S. Geological Survey). Equation: (Band 4-Red/Band 2-Blue)	[41]
Age (A)	GEOM	Geomorphons	30 m	Copernicus GLO-30 (DSM) and ANADEM (DTM). SAGA GIS 9.3.1	[42]

For the organism factor, the Soil-Adjusted Vegetation Index (SAVI) [43] and daytime Land Surface Temperature (LST) [44] were used. These were obtained from LANDSAT-8 OLI multispectral imagery and the MODIS/061/MOD11A1 daily time series (2000–2024) via GEE [45]. In the Amazon biome, organic matter inputs and high temperatures strongly influence particle aggregation, darkening of the surface horizon, water infiltration, soil erodibility, nutrient dynamics, and forest biomass [32].

Terrain attributes were derived using two elevation models: the Digital Surface Model (DSM) and Digital Terrain Model (DTM). The DSM was obtained from the Copernicus DEM GLO-30 (1 Arc Second), which provides global coverage at 30 m spatial resolution, selected for its free availability, high quality, and proven accuracy in tropical environments [46].

The DTM was derived from the South America Digital Terrain Model (ANADEM) [47], also at 30 m resolution. The development of ANADEM involved (i) calculating the Enhanced Vegetation Index (EVI), Normalized Difference Moisture Index (NDMI), Modified Simple Ratio (MSR), and Modified Soil-Adjusted Vegetation Index (MSAVI) from multispectral imagery acquired by the LANDSAT- 8 Operational Land Imager (OLI) and Sentinel-2 MultiSpectral Instrument (MSI); (ii) incorporating vegetation height data from the Global Ecosystem Dynamics Investigation (GEDI), collection L2A; (iii) using the DSM from Copernicus GLO-30; (iv) applying the TreeBoost machine learning algorithm to estimate bias induced by vegetation; and (v) obtaining the DTM by subtracting the vegetation-induced bias from the Copernicus DSM, accounting for correction based on the vegetation fractional cover.

Terrain covariates were derived from the DSM and DTM using the Basic Terrain Analysis module in SAGA GIS v. 9.3.1 [48], resampled to the UTM SIRGAS 2000 Zone 23S projection at 30 m resolution. These include CNBL, DD, MRRTF, MRVBF, PFC, RSP, and GEOM. Linear drainage density (DD) was calculated in ArcGIS 10.5 as the ratio of total channel length to basin area [49], according to [50], who reported its significant contribution to model performance and soil map extrapolation. CNBL, DD, MRRTF, MRVBF, PFC, and RSP represent terrain surface configuration and relate to the spatial distribution of soils across landscape scales [32]. GEOM captures terrain form patterns, allowing for intuitive classification of topographic features and providing useful estimates of surface age within the SCORPAN framework [7,42].

Information regarding geology and Iron Oxide Ratio (IOR) was sourced from Brazilian Institute of Geography and Statistics (IBGE) and LANDSAT-8 multispectral imagery, respectively. Parent material influences key soil properties, including consolidation, texture, and composition. The accumulation of iron oxides and hydroxides such as hematite and goethite occur in highly weathered soils and provides an estimate of the degree of soil weathering [32].

2.4. Machine Learning Algorithms

Machine learning algorithms are capable of modeling nonlinear relationships between SCORPAN covariates, including soil classes as targets and terrain attributes, location, and temporal series as predictor variables, which can not only improve mapping accuracy but also quantify uncertainty. In this study, four tree-based algorithms were used—Random Forest (RF), Fast implementation of Random Forests (Ranger), XGBoost (Xgb), and C5.0 (C5)—as well as an ensemble learning (EL) approach combining these algorithms (Table 3). Testing different algorithms is essential for identifying the most suitable method for local conditions [1].

The RF [51] and Ranger [52] algorithms share the following fundamental principles: (i) bagging (Bootstrap Aggregating), where each decision tree is trained on a randomly selected sample with replacement from the training set; and (ii) feature bagging, in which only a subset of features is selected for each decision tree in the forest. Final predictions are made by majority vote, as in the current classification study. Differences between the algorithms include computational efficiency, with Ranger being faster, and the use of different R packages—“randomForest” and “ranger” for RF and Ranger, respectively.

These models have three hyperparameters for optimization: (i) “ntree” (1000) for RF and “num.trees” (1000) for Ranger; (ii) “nodesize” (5); and (iii) “mtry” (set to 10). Among these, “mtry” was optimized using the “caret” package, testing 10 different values according to the number of covariates, with the value yielding the best performance selected, in accordance with [1].

XGBoost (Xgb), developed by [53], is based on gradient-boosted decision trees. In this model, new trees are created to correct the errors of the previous tree, and model performance is assessed based on the difference between predicted and observed values. Xgb offers advantages such as improved predictive performance via gradient descent, enhanced model accuracy, computational efficiency, parallel and distributed computing, and the ability to handle missing values, making it well-suited for complex prediction problems. The algorithm was implemented using the “xgbTree” package, with the hyperparameter “nrounds” set to 1000, while other hyperparameters—“max_depth,” “eta,” “gamma,” “colsample_bytree,” “min_child_weight,” and “subsample”—were optimized using the “caret” package.

C5.0, proposed by [54], is based on a broad decision tree capable of handling noise and missing data. The algorithm generates classifiers that help reduce the cost of misclassification rather than simply minimizing the error rate [55]. The C5.0 model includes three hyperparameters: (i) “trials” (100), (ii) “model” (tree), and (iii) “winnow” (false). The final model was implemented using the “C50” package.

In PDSM, a single machine learning algorithm is often used to implement the model. However, ensemble learning, which has the advantage of combining the strengths of different algorithms, remains underutilized. Ensemble learning encompasses three main techniques: bagging, boosting, and stacking. In this study, the bagging technique proposed by [51] was employed, aiming to combine similar models (e.g., tree-based) to reduce variance and increase model robustness. Final predictions can be aggregated by majority vote (hard voting) or probability averaging (soft voting), with the latter approach applied in the present study [56].

As the target, the dataset of 270 samples for both LD1 and LD2 (representing the soil taxonomic levels of Order and Great Group, respectively) was split into 80% for training with 5-fold cross-validation and 20% for testing. Using the “caret” package, this split was performed with the “createDataPartition” function, which executes more representative random sampling [1]. Subsequently, to identify the optimal subset of predictor covariates for model performance, Recursive Feature Elimination (RFE) was applied. With the adjusted parameters, 100 iterations were run, with each generated map resulting from an independent sampling iteration. The best model was selected based on the highest performance according to the Kappa index and the importance of the respective predictor covariates in predicting the targets (Figure 4).

The parameters pre-configured before the start of the process are called hyperparameters. Their selection can drastically change the model’s performance, and is essential for the reproducibility of the modeling. Therefore, in this study, we used the configurations shown in

Table 4. Hyperparameters used in modeling according to the algorithms, at different levels details different (LD1 and 2) using DSM and DTM.

Algorithm	LD	Hyperparameters
		DSM	DTM
RF	LD1	mtry = 2; ntree = 1000	mtry = 2; ntree = 1000
	LD2	mtry = 4, ntree = 1000	mtry = 6; ntree = 1000
Ranger	LD1	mtry = 10; num.trees = 1000	mtry = 4; num.trees = 1000
	LD2	mtry = 2; num.trees = 1000	mtry = 2; num.trees = 1000
Xgb	LD1	nrounds = 1000; max_deph = 6; eta = 0.01; gamma = 0; colsample_bytree = 0.8; min_child_weight = 1; subsample = 0.8
	LD2	nrounds = 1000; max_deph = 3; eta = 0.01; gamma = 0; colsample_bytree = 0.8; min_child_weight = 1; subsample = 0.8
C5	LD1	trials = 100. Multiple model combination (10 models C 5.0)
	LD2
EL	LD1	bestTune
	LD2

.

To represent the models used, we employed the following structure: abbreviation of the algorithm, e.g., RF, plus the origin of the terrain attributes used as environmental covariates, either DSM (S) or DTM (T), and the level of detail of the taxonomic units and associations, either LD1 or LD2. Thus, the nomenclature of the XgbS2 model, for example, means that the Xgboost algorithm was used, with terrain covariates derived from DSM, at LD2.

In this study we used Google Gemini (Free Version) to assist in refining the English translation and performing grammar checks of the manuscript text, and generating the graphical abstract figure. The results from the generative artificial intelligence (GenAI) tool were used to assist only and were carefully reviewed, edited and validated by the authors. All methodological decisions, data processing, statistical analyses, interpretation of results, and scientific conclusions were performed exclusively by the authors, who take full responsibility for the content of the manuscript.

2.5. Spatial Prediction Performance on the Test Dataset

The performance of the algorithms was evaluated using the mean Kappa index (k) and overall accuracy. The Kappa index (Equation (1)) indicates the level of agreement in predicting the mapping unit [57].

$$k={\displaystyle \frac{n{\sum }_{i=1}^C{n}_{ii}-{\sum }_{i=1}^C{n}_{i+}+n_{+i}}{n^{2 }-{\sum }_{i=1}^C{n}_{i+}+n_{+i}}}$$

(1)

where $n_{ii}$ is the value in row i and column $i$; $n_{i+}$ is the sum of row $i$ and $n_{+i}$ is the sum of column $i$ in the confusion matrix; $n$ is the total number of samples; and $C$ is the total number of classes.

The k was interpreted using the scale proposed by [58], which defines the following ranges: Zero (0)—“Terrible,” meaning no agreement beyond chance; “Bad” (0.01–0.20); “Fair” (0.21–0.40); “Moderate” (0.41–0.60); “Very Good” (0.61–0.80); and “Excellent” (0.81–1), indicating almost perfect agreement between the models on the test data.

The Accuracy and k results from the n = 100 maps generated for each algorithm were compared using Tukey’s Honest Significant Difference (HSD) test at a 1% confidence level to assess statistically significant differences among the algorithms studied, including the ensemble learning approach (Equation (2)).

$$HSD=q\sqrt{{\displaystyle \frac{MSE}{n_c}}}$$

(2)

where $HSD$ is the minimum difference between two group means required for statistical significance; $q$ is the critical value from the table; $MSE$ is the mean squared error; and $n_c$ is the sample size for each group.

Overall accuracy, defined as the degree of agreement between model predictions and observed values, was calculated as the number of correctly classified cases divided by the total number of predictions [59], according to Equation (3).

$$Accuracy={\displaystyle \frac{{\sum }_{i=1}^C{TP}_i}{N}}$$

(3)

where $TP$ refers to the correctly classified instances of class $i$ and $C$ is the total number of classes, divided by the total number of cases $N$.

2.6. Pixel-by-Pixel Spatial Prediction Performance on the Legacy Map

The digital maps produced by the tested algorithms were compared pixel-by-pixel with the legacy soil map using the Semi-Automatic Classification Plugin (SCP) [60] implemented in QGIS Desktop 3.34.8. Model performance was evaluated through a confusion matrix, User’s accuracy (UA), Producer’s accuracy (PA), overall accuracy (OA), and Kappa index $\left(\widehat{k}\right)$ [61].

Overall accuracy (OA) measures the proportion of correctly classified pixels relative to the reference data (Equation (4)).

$$OA={\displaystyle \frac{\sum _{i=1}^k{n}_{ii}}{N}}$$

(4)

where $n_{ii}$ is the number of correctly classified samples for class $i$, corresponding to the diagonal of the confusion matrix; $k$ is the total number of classes; and $N$ is the total number of samples (sum of all elements in the matrix).

To properly assess UA and PA, the F1 score (Equation (5)) was calculated, which represents the harmonic mean of these two metrics. The F1 score was computed individually for each mapping unit class and ranges from 0 to 1, with values closer to 1 indicating higher predictive accuracy [62].

$$F1 score=2\left({\displaystyle \frac{UA\times PA}{UA+PA}}\right)$$

(5)

where $UA$ is the percentage of samples correctly predicted for a given class relative to the total number of samples predicted for that class, also known as precision, and $PA$ (or recall) is the proportion of samples correctly predicted for a given class.

2.7. Uncertainty of Algorithms Based on Soil Mapping Unit Classification

To estimate the prediction error of the generated digital soil maps—i.e., how close the predicted values are to the true values, also referred to as uncertainty—the probabilistic prediction of class occurrence for each mapping unit (MU) was calculated using the number of votes from the n = 100 maps generated during algorithm execution. This assessment, however, is only of practical value to the end user if it is accurate and reliable [63].

The confusion index (CI), which quantifies the uncertainty between dominant and subdominant soil classes in the mapping units (MUs) (Equation (6)) [33], and the normalized Shannon entropy [64], which measures the spatial prediction uncertainty (Equation (7)), are defined as follows:

$$CI=[1-\left({\mu }_{max i}-{\mu }_{(\max - 1)i}\right)]$$

(6)

where ${\mu }_{maxi}$ is the probability of the most likely class for soil sampling point $i,$ and ${\mu }_{(\max -1)i}$ is the probability of the second most likely class for soil sampling point $i$. CI values range from 0 to 1, with higher values indicating greater uncertainty.

$$Entropy={\displaystyle \frac{-{\sum }_{k=1}^{n}pk·{log}_2\left(pk\right)}{{log}_2\left(n\right)}}$$

(7)

where $pk$ is the probability of class $k$ and $n$ is the total number of soil classes in the mapping units (MUs). The entropy values were normalized, where 0 indicates maximum confidence in the predicted class and 1 indicates maximum uncertainty in the predictions.

3. Results

3.1. Model Performance

We compared the performance of the models in predicting the MUs in LD1 and LD2 (Figure 5a–d). All models presented a “very good” k (0.61–0.80), except for models C5S1 and C5T1 (LD1) and ELT2 (LD2), which obtained a “moderate” k (0.41–0.60), according to the scale proposed by [58].

In LD1, the ELS1, RangerT1, RFT1, ELT1, and RangerS1 models showed the highest average accuracy and k values, ranging from 0.74 to 0.72 and 0.67 to 0.63, respectively. In LD2, only the ELS2 model showed a higher average than the others, with 0.75 and 0.72 for accuracy and k, respectively. On the other hand, the C5 algorithms showed the lowest performance in both LD1 and LD2.

We evaluated the weight of the 15 environmental covariates in all models studied and in the prediction of soil mapping units in LD1 and LD2 (Figure 6a,b). When considering the ELS1 and ELS2 models, the environmental covariates Tmean_22k, CNBL, Altitude, and LST (ELS1) and Tmean_22k, CNBL, LST, and Altitude (ELS2) were the ones that contributed most to the models.

3.2. Internal Validation

We compared the predicted map with the legacy (reference) map, pixel-by-pixel, based on the boundaries of the predicted taxonomic units. The OA and $\widehat{k}$ values (

Table 5. Overall accuracy (OA) and Kappa index ($\widehat{k}$) for the models at two levels of detail (LD1 and LD2).

Levels of Detail
LD1			LD2
Models	OA	$\widehat{k}$	Models	OA	$\widehat{k}$
RFS1	0.72	0.63	RFS2	0.71	0.66
RangerS1	0.72	0.63	RangerS2	0.71	0.66
XgbS1	0.72	0.62	XgbS2	0.69	0.64
C5S1	0.68	0.58	C5S2	0.70	0.64
ELS1	0.69	0.59	ELS2	0.68	0.62
RFT1	0.71	0.62	RFT2	0.72	0.67
RangerT1	0.72	0.63	RangerT2	0.71	0.66
XgbT1	0.72	0.63	XgbT2	0.70	0.65
C5T1	0.71	0.61	C5T2	0.69	0.63
ELT1	0.71	0.62	ELT2	0.69	0.63

) showed very little variation when comparing all of the models tested, in LD1 and LD2. Despite the reduction in the number of soil mapping units from nine in LD2 to five in LD1, we observed that the average $\widehat{k}$ decreased from 0.65 to 0.62.

In Figure 7, we observe that the RangerT1, RangerS1, and RFS1 models in LD1 and the RangerS2, RangerT2, and RF2 models in LD2 presented the highest absolute values of PA and UA (Figure 7a,b). We observed that RQg5 and GXve2 presented the lowest PA metrics, frequently below 50% in several models, highlighting difficulties in spatial discrimination.

GZn1 showed the highest UA in almost all models—RangerS2, RangerT2, RFS2, RFT2, and C5S2—reaching 96%. The PA was also consistently high, reaching 89% in the C5S2 model. It is also worth highlighting that the high F1-score in GZn1 revealed that the climate (Tmean_22k) and topography (CNBL) factors were the most important predictive covariates in predicting this unit, which indicates soils in hydromorphic environments.

3.3. Uncertainty of Spatial Predictions

In Figure 8 and Figure 9 (entropy maps) and Figure 10 and Figure 11 (CI maps), we observe that the ELS and ELT models for both LD1 and LD2 similarly presented a larger area with predicted MUs with less uncertainty (entropy and CI close to 0) when compared to other traditional models.

The spatial uncertainty of the predictive models was further evaluated for soil mapping units under hydromorphic conditions, which are characteristic of the carbon-rich valley bottoms and alluvial plains of the Amazon. A comparative analysis of the entropy and confusion index (CI) across mapping units LD1 and LD2 is illustrated in Figure 12 and Figure 13.

4. Discussion

4.1. Evaluation of Modeling Metrics

In LD1, the stability in the prediction (Accuracy and k) of the MUs of the ELS1, RangerT1, ELT1, RFT1, and RangerS1 models was statistically higher and equal to that of the C5S1 and C5T1 models. In LD2, the stability in the prediction of the ELS2 model was statistically superior to all individual predictive models, including the ELT2 model (Figure 5).

The greater complexity of the spatial distribution of the MUs in LD2 may be related to the difference between the larger number of predictive models with higher precision and k values in LD1 compared to LD2 (Figure 5). This may be related to the potential of the EL approach to mitigate errors, mainly by using a conventional soil map as a reference. These results are consistent with those of [65,66]

Unlike our hypothesis, in general, the use of terrain attributes derived from ANADEM (DTM) and the aggregation of soil units in LD1 did not improve performance when used in modeling. This result demonstrated that the use of environmental covariates derived from DSM was sufficient in predicting soil units. The aggregation of soil units from a legacy map did not prove to be a strategy capable of improving model performance, as there was a 5-percentage point decrease in the k of ELS2 compared to ELS1 (Figure 5). This proves that despite the challenges regarding the high uncertainty implicit in the legacy map, these provide valuable information on the spatial distribution of soils [67], particularly in the Amazon region.

These results demonstrate the ability of EL to aggregate the strengths of individual algorithms. There was a 4-percentage-point difference in the k between ELS1 (0.67) and ELS2 (0.72), showing that ELS2 was more balanced than ELS1 (Figure 5), despite disagreeing with the results of [50], which justified the reduction in model performance due to the increased complexity of the MUs. In our study, the reduction in ELS1 performance compared to ELS2 may be related to error amplification by aggregating LD2 MUs to LD1 from the conventional map (reference).

The small differences in accuracy and k (Figure 5), especially in the models EL, Ranger, and RF in LD1, may be related to the fact that these algorithms all use a bagging structure, while Xgb uses a boosting structure [26,68,69].

The potential improvement in k values achieved by EL compared to traditional models for soil class prediction has been poorly reported in the literature [67]. Our results are consistent with those of [65,66], who also reported superior performance of ensemble models compared to traditional approaches. These results indicate that ELS mitigated the spatial prediction errors of individual algorithms, mainly when using a conventional soil map as a reference.

Our findings showed that paleoclimate, represented by the covariate Tmean_22k (Figure 6), was more important than the covariates representing the current climate. Similar results were found by [6], who observed the estimated precipitation from 22,000 years ago in the south of the state of Minas Gerais. During this period, it is believed that, in Brazil, a warmer and rainier climate favored erosive processes related to pedogenesis. Even in gentler relief conditions, both water infiltration into the soil profile and surface runoff can accelerate weathering [6]. Particularly in the development area of our study, characterized by its proximity to the Atlantic Ocean, this may justify the distribution of the MUs in the area.

The terrain attribute CNBL was one of the most important covariates in the ELS1 and ELS2 models (Figure 6); as a covariate derived from the drainage network, it is frequently associated with soil erosion and sedimentation and plays a key role in soil–landscape stratification [50,69,70]. The CNBL calculates the vertical distance of drainage channels, capable of separating soils under hydromorphic influence from soils in plateau areas [71]. In our work, approximately 40% of the mulches are under hydromorphic influence, which justifies this covariate being among the most important. Based on these results, we observed that the selection of uncorrelated covariates that encompass all SCORPAN factors may be related to the higher Kappa indices and precision.

These findings support the potential integration of tacit knowledge with machine learning. However, this should be approached with caution, since pedological knowledge, in addition to being reflected in the selection of relevant covariates, requires subsequent interpretation and analysis of the resulting patterns, both from a spatial prediction and soil science perspective [10,72].

We compared all models pixel by pixel with the conventional map (Table 4) in terms of the relationship between the reference and predicted MUs and by means of a confusion matrix (

Table A1. Pixel-by-pixel internal validation confusion matrix between the predicted map generated by the RFS1 and RFT1 models and the legacy map (reference), in LD1.

RFS1	Reference							RFT1	Reference
LD a	R	G	E	L	P	Total	UA %	LD a	R	G	E	L	P	Total	UA %
R	188,925	41,257	5626	72,208	22,277	330,293	57	R	193,479	33,912	4569	95,203	16,450	343,613	56
G	19,300	194,518	3019	0	662	217,499	89	G	24,428	205,370	6183	0	1844	237,825	86
E	9585	12,833	41,022	0	0	63,440	65	E	9147	10,645	38,915	0	0	58,707	66
L	33,780	0	0	149,874	5187	188,841	79	L	23,589	0	0	127,329	5272	156,190	82
P	21,924	227	0	10,971	84,462	117,584	72	P	22,844	281	0	10,510	89,017	122,652	73
Total	273,514	248,835	49,667	233,053	112,588	917,657		Total	273,487	250,208	49,667	233,042	112,583	918,987
PA %	69	79	82	64	75			PA %	71	83	78	55	79
	OA = 0.72	Kappa = 0.63							OA = 0.71	Kappa = 0.62

LD ^a: Soil classification according to SiBCS [28]. UA: User’s accuracy; PA: Producer’s accuracy; OA: overall accuracy.

,

Table A2. Pixel-by-pixel internal validation confusion matrix between the predicted map generated by the RangerS1 and RangerT1 models and the legacy map (reference), in LD1.

RangerS1	Reference							RangerT1	Reference
LD a	R	G	E	L	P	Total	UA %	LD a	R	G	E	L	P	Total	UA %
R	191,271	38,602	5786	72,568	24,527	332,754	57	R	198,735	31,733	4530	95,211	18,447	348,656	57
G	20,793	196,551	2830	0	736	220,910	89	G	25,005	208,141	6067	0	2037	241,250	86
E	9767	13,566	41,051	0	0	64,384	64	E	8444	10,270	39,070	0	0	57,784	68
L	37,883	0	0	154,306	10,661	202,850	76	L	24,610	0	0	129,892	6670	161,172	81
P	13,800	116	0	6179	76,664	96,759	79	P	16,693	64	0	7939	85,429	110,125	78
Total	273,514	248,835	49,667	233,053	112,588	917,657		Total	273,487	250,208	49,667	233,042	112,583	918,987
PA %	70	80	82	66	68			PA %	72	84	78	56	76
	OA = 0.72	Kappa = 0.63							OA = 0.72	Kappa = 0.63

LD ^a: Soil classification according to SiBCS [28]. UA: User’s accuracy; PA: Producer’s accuracy; OA: overall accuracy.

,

Table A3. Pixel-by-pixel internal validation confusion matrix between the predicted map generated by the XgbS1 and XgbT1 models and the legacy map (reference), in LD1.

XgbS1	Reference							XgbT1	Reference
LD a	R	G	E	L	P	Total	UA %	LD a	R	G	E	L	P	Total	UA %
R	183,473	42,970	5049	65,142	21,162	317,796	58	R	191,145	35,753	4883	77,416	21,197	330,394	58
G	23,687	193,069	4355	222	1152	222,485	87	G	24,807	203,672	6193	0	2436	237,108	86
E	8091	12,563	40,263	0	0	60,917	66	E	8009	10,570	38,591	0	0	57,170	68
L	35,862	3	0	155,065	6706	197,636	78	L	29,316	0	0	142,863	6069	178,248	80
P	22,401	230	0	12,624	83,568	118,823	70	P	20,210	213	0	12,763	82,881	116,067	71
Total	273,514	248,835	49,667	233,053	112,588	917,657		Total	273,487	250,208	49,667	233,042	112,583	918,987
PA %	67	78	81	66	74			PA %	70	82	77	61	74
	OA = 0.72	Kappa = 0.62							OA = 0.72	Kappa = 0.63

LD ^a: Soil classification according to SiBCS [28]. UA: User’s accuracy; PA: Producer’s accuracy; OA: overall accuracy.

,

Table A4. Pixel-by-pixel internal validation confusion matrix between the predicted map generated by the C5S1 and C5T1 models and the legacy map (reference), in LD1.

C5S1	Reference							C5T1	Reference
LD a	R	G	E	L	P	Total	UA %	LD a	R	G	E	L	P	Total	UA %
R	170,828	46,142	5554	74,504	13,460	310,488	55	R	195,243	31,002	4360	99,585	15,934	346,124	56
G	32,521	186,297	2601	1507	7008	229,934	81	G	27,938	209,649	7074	61	2788	247,510	85
E	11,649	16,126	41,512	0	226	69,513	60	E	6927	9507	38,233	0	60	54,727	70
L	38,834	115	0	145,647	11,675	196,271	74	L	24,723	0	0	121,181	8676	154,580	78
P	19,682	155	0	11,395	80,219	111,451	72	P	18,656	50	0	12,215	85,125	116,046	73
Total	273,514	248,835	49,667	233,053	112,588	917,657		Total	273,487	250,208	49,667	233,042	112,583	918,987
PA %	62	75	83	62	71			PA %	71	84	77	52	76
	OA = 0.68	Kappa = 0.58							OA = 0.71	Kappa = 0.61

LD ^a: Soil classification according to SiBCS [28]. UA: User’s accuracy; PA: Producer’s accuracy; OA: overall accuracy.

,

Table A5. Pixel-by-pixel internal validation confusion matrix between the predicted map generated by the ELS1 and ELT1 models and the legacy map (reference), in LD1.

ELS1	Reference							ELT1	Reference
LD a	R	G	E	L	P	Total	UA %	LD a	R	G	E	L	P	Total	UA %
R	136,555	31,631	2981	32,173	14,456	217,796	63	R	151,403	20,900	4604	57,041	15,889	249,837	61
G	28,668	196,765	2921	637	3771	232,762	85	G	41,468	215,041	8094	102	2554	267,259	80
E	18,405	17,064	43,374	0	0	78,843	55	E	8163	12,920	36,930	0	0	58,013	64
L	58,218	408	1	189,792	29,048	277,467	68	L	47,344	93	1	165,135	12,029	224,602	74
P	19,533	382	0	5083	61,382	86,380	71	P	23,187	392	0	10,116	81,449	115,144	71
Total	261,379	246,250	49,277	227,685	108,657	893,248		Total	271,565	249,346	49,629	232,394	111,921	914,855
PA %	52	76	91	86	37			PA %	55	87	74	71	73
	OA = 0.69	Kappa = 0.59							OA = 0.71	Kappa = 0.62

LD ^a: Soil classification according to SiBCS [28]. UA: User’s accuracy; PA: Producer’s accuracy; OA: overall accuracy.

,

Table A6. Pixel-by-pixel internal validation confusion matrix between the predicted map generated by the RFS2 model and the legacy map (reference), in LD2.

RFS2	Reference
LD a	RQo1	RQo5	RQg5	RQg1	GXve2	ESo	LAd	PAd	GZn1	Total	UA %
RQo1	54,138	3556	0	194	4512	3477	122	0	24,575	90,574	60
RQo5	340	65,847	0	3655	8825	566	6428	9116	1735	96,512	68
RQg5	0	0	1132	0	0	0	0	0	1254	2386	47
RQg1	0	4639	0	26,567	6	4	31,941	2206	2	65,365	41
GXve2	1881	20,538	0	1129	30,767	632	71	3340	3945	62,303	49
ESo	8655	2985	0	0	13,197	43,817	0	0	3266	71,920	61
LAd	0	17,427	0	31,391	6	0	183,408	6050	0	238,282	77
PAd	0	6307	0	22,069	309	0	11,083	91,888	0	131,656	70
GZn1	594	0	458	0	4474	1171	0	0	151,962	158,659	96
Total	65,608	121,299	1590	85,005	62,096	49,667	233,053	112,600	186,739	917,657
PA %	82	54	72	31	49	88	79	82	82
	OA = 0.71	Kappa = 0.66

LD ^a: Soil classification according to SiBCS [28]. UA: User’s accuracy; PA: Producer’s accuracy; OA: overall accuracy.

,

Table A7. Pixel-by-pixel internal validation confusion matrix between the predicted map generated by the RFT2 model and the legacy map (reference), in LD2.

RFT2	Reference
LD a	RQo1	RQo5	RQg5	RQg1	GXve2	ESo	LAd	PAd	GZn1	Total	UA %
RQo1	53,149	4705	0	0	5307	3321	0	0	15,898	82,380	65
RQo5	30	66,391	0	2094	6583	620	16,728	3814	175	96,435	69
RQg5	0	0	1124	0	0	0	0	0	2572	3696	30
RQg1	0	3395	0	32,852	16	0	36,161	2279	0	74,703	44
GXve2	2284	24,644	0	1131	33,417	871	101	5554	2290	70,292	48
ESo	9109	2857	0	0	12,147	42,694	0	0	5230	72,037	59
LAd	0	12,225	0	28,539	0	0	171,097	6474	0	218,335	78
PAd	0	7039	0	20,381	354	0	8955	94,473	0	131,202	72
GZn1	1038	0	489	0	4278	2161	0	0	161,941	169,907	95
Total	65,610	121,256	1613	84,997	62,102	49,667	233,042	112,594	188,106	918,987
PA %	81	55	70	39	54	86	73	84	87
	OA = 0.72	Kappa = 0.67

LD ^a: Soil classification according to SiBCS [28]. UA: User’s accuracy; PA: Producer’s accuracy; OA: overall accuracy.

,

Table A8. Pixel-by-pixel internal validation confusion matrix between the predicted map generated by the RangerS2 model and the legacy map (reference), in LD2.

RangerS2	Reference
LD a	RQo1	RQo5	RQg5	RQg1	GXve2	ESo	LAd	PAd	GZn1	Total	UA %
RQo1	53,729	4683	0	749	4843	3552	786	1	19,137	87,480	61
RQo5	54	66,020	0	3822	7695	603	11,993	9103	77	99,367	66
RQg5	0	0	1107	0	0	0	0	0	1394	2501	44
RQg1	0	2344	0	28,351	5	0	33,080	2180	0	65,960	43
GXve2	1973	21,940	0	1044	31,541	524	32	3343	3576	63,973	49
ESo	8802	2833	0	0	13,236	43,229	0	0	2982	71,082	61
LAd	0	17,862	0	32,384	14	0	178,797	9860	0	238,917	75
PAd	0	5613	0	18,598	204	0	8352	88,113	0	120,880	73
GZn1	1050	4	483	57	4558	1759	13	0	159,573	167,497	95
Total	65,608	121,299	1590	85,005	62,096	49,667	233,053	112,600	186,739	917,657
PA %	82	54	71	33	51	87	77	78	86
	OA = 0.71	Kappa = 0.66

LD ^a: Soil classification according to SiBCS [28]. UA: User’s accuracy; PA: Producer’s accuracy; OA: overall accuracy.

,

Table A9. Pixel-by-pixel internal validation confusion matrix between the predicted map generated by the RangerT2 model and the legacy map (reference), in LD2.

RangerT2	Reference
LD a	RQo1	RQo5	RQg5	RQg1	GXve2	ESo	LAd	PAd	GZn1	Total	UA %
RQo1	53,615	5819	0	0	5664	3541	0	0	14,057	82,696	65
RQo5	0	66,310	0	1987	6075	665	21,046	4319	2	100,404	66
RQg5	0	0	1131	0	0	0	0	0	2497	3628	31
RQg1	0	2742	0	36,141	3	0	38,462	3636	0	80,984	45
GXve2	2161	27,066	0	1049	33,881	782	97	6083	1809	72,928	46
ESo	7508	2811	0	0	11,580	42,173	0	0	4060	68,132	62
LAd	0	11,067	0	29,119	0	0	167,273	10,527	0	217,986	77
PAd	0	5440	0	16,701	268	0	6164	88,029	0	116,602	75
GZn1	2326	1	482	0	4631	2506	0	0	165,681	175,627	94
Total	65,610	121,256	1613	84,997	62,102	49,667	233,042	112,594	188,106	918,987
PA %	82	55	70	43	54	85	72	78	89
	OA = 0.71	Kappa = 0.66

LD ^a: Soil classification according to SiBCS [28]. UA: User’s accuracy; PA: Producer’s accuracy; OA: overall accuracy.

,

Table A10. Pixel-by-pixel internal validation confusion matrix between the predicted map generated by the XgbS2 model and the legacy map (reference), in LD2.

XgbS2	Reference
LD a	RQo1	RQo5	RQg5	RQg1	GXve2	ESo	LAd	PAd	GZn1	Total	UA %
RQo1	53,761	2446	0	1829	4349	2599	4942	0	23,271	93,197	58
RQo5	599	66,024	0	2461	8977	573	13,897	8119	566	101,216	65
RQg5	45	0	1061	0	0	0	0	0	1476	2582	41
RQg1	0	1897	0	28,443	102	0	30,407	4527	0	65,376	44
GXve2	3216	23,921	0	1479	31,883	2476	20	3355	6051	72,401	44
ESo	7093	2359	0	0	11,325	42,870	0	0	2867	66,514	64
LAd	158	17,913	0	30,212	66	0	170,003	7728	18	226,098	75
PAd	0	6739	0	20,581	416	0	13,784	88,871	0	130,391	68
GZn1	736	0	529	0	4978	1149	0	0	152,490	159,882	95
Total	65,608	121,299	1590	85,005	62,096	49,667	233,053	112,600	186,739	917,657
PA %	82	54	68	33	51	86	73	79	82
	OA = 0.69	Kappa = 0.64

LD ^a: Soil classification according to SiBCS [28]. UA: User’s accuracy; PA: Producer’s accuracy; OA: overall accuracy.

,

Table A11. Pixel-by-pixel internal validation confusion matrix between the predicted map generated by the XgbT2 model and the legacy map (reference), in LD2.

XgbT2	Reference
LD a	RQo1	RQo5	RQg5	RQg1	GXve2	ESo	LAd	PAd	GZn1	Total	UA %
RQo1	53,359	2086	0	91	4572	3195	0	0	19,110	82,413	65
RQo5	929	65,839	0	2683	7391	868	17,463	4910	1309	101,392	65
RQg5	47	0	1136	0	0	0	0	0	3807	4990	23
RQg1	0	2931	0	36,139	61	0	37,090	6726	0	82,947	44
GXve2	3002	28,726	0	821	35,478	2183	18	6038	5232	81,498	44
ESo	7350	2900	0	0	9738	41,553	0	0	4041	65,582	63
LAd	62	12,237	0	26,556	64	1	166,412	7024	33	212,389	78
PAd	0	6537	0	18,707	492	0	12,059	87,896	0	125,691	70
GZn1	861	0	477	0	4306	1867	0	0	154,574	162,085	95
Total	65,610	121,256	1613	84,997	62,102	49,667	233,042	112,594	188,106	918,987
PA %	81	54	70	43	57	83	71	78	83
	OA = 0.70	Kappa = 0.65

LD ^a: Soil classification according to SiBCS [28]. UA: User’s accuracy; PA: Producer’s accuracy; OA: overall accuracy.

,

Table A12. Pixel-by-pixel internal validation confusion matrix between the predicted map generated by the C5S2 model and the legacy map (reference), in LD2.

C5S2	Reference
LD a	RQo1	RQo5	RQg5	RQg1	GXve2	ESo	LAd	PAd	GZn1	Total	UA %
RQo1	51,696	2738	0	9	4875	3992	42	0	11,932	75,284	69
RQo5	902	72,154	0	5445	10,809	1528	27,894	8641	487	127,860	56
RQg5	0	0	1135	0	0	0	0	0	1702	2837	40
RQg1	0	1257	0	25,207	9	0	30,611	2208	0	59,292	43
GXve2	1536	18,622	0	1426	29,497	545	62	5285	3198	60,171	49
ESo	7453	3744	0	0	12,321	42,640	0	0	3189	69,347	61
LAd	0	18,048	0	32,769	4	0	162,657	9284	0	222,762	73
PAd	0	4736	0	20,148	122	0	11,779	87,143	0	123,928	70
GZn1	4021	0	455	0	4459	962	0	0	166,231	176,128	94
Total	65,608	121,299	1590	85,004	62,096	49,667	233,045	112,561	186,739	917,609
PA %	79	59	73	30	47	86	70	77	89
	OA = 0.70	Kappa = 0.64

LD ^a: Soil classification according to SiBCS [28]. UA: User’s accuracy; PA: Producer’s accuracy; OA: overall accuracy.

,

Table A13. Pixel-by-pixel internal validation confusion matrix between the predicted map generated by the C5T2 model and the legacy map (reference), in LD2.

C5T2	Reference
LD a	RQo1	RQo5	RQg5	RQg1	GXve2	ESo	LAd	PAd	GZn1	Total	UA %
RQo1	49,802	2809	0	0	4610	3408	0	0	8925	69,554	72
RQo5	622	72,464	0	3700	10,022	1688	27,088	3923	108	119,615	61
RQg5	0	0	1284	0	0	0	0	0	3718	5002	26
RQg1	116	2110	0	34,514	68	0	46,856	2851	1444	87,959	39
GXve2	2288	22,894	0	1175	32,027	1349	131	7915	3209	70,988	45
ESo	6982	3794	0	0	11,533	41,273	0	0	4047	67,629	61
LAd	12	13,031	0	26,214	0	0	148,477	12,780	0	200,514	74
PAd	0	4154	0	19,394	101	0	10,490	85,125	2	119,266	71
GZn1	5788	0	329	0	3741	1949	0	0	166,653	178,460	93
Total	65,610	121,256	1613	84,997	62,102	49,667	233,042	112,594	188,106	918,987
PA %	76	60	80	41	51	83	64	76	89
	OA = 0.69	Kappa = 0.63

LD ^a: Soil classification according to SiBCS [28]. UA: User’s accuracy; PA: Producer’s accuracy; OA: overall accuracy.

,

Table A14. Pixel-by-pixel internal validation confusion matrix between the predicted map generated by the ELS2 model and the legacy map (reference), in LD2.

ELS2	Reference
LD a	RQo1	RQo5	RQg5	RQg1	GXve2	ESo	LAd	PAd	GZn1	Total	UA %
RQo1	51,727	4349	0	451	4749	2036	1434	0	17,950	82,696	63
RQo5	167	55,305	0	3067	7181	417	19,355	8142	343	93,977	59
RQg5	0	0	632	0	0	0	0	0	1108	1740	36
RQg1	0	2760	0	29,751	126	0	26,572	8280	0	67,489	44
GXve2	2093	27,131	0	1724	28,767	1864	95	5685	1979	69,338	41
ESo	10,538	3232	3	0	15,315	43,106	0	0	5511	77,705	55
LAd	7	15,648	0	28,836	33	1	166,439	12,817	11	223,792	74
PAd	0	5112	0	16,945	292	0	13,546	73,651	0	109,546	67
GZn1	896	113	595	87	5030	1799	228	0	157,840	166,588	95
Total	65,428	113,650	1230	80,861	61,493	49,223	227,669	108,575	184,742	892,871
PA %	79	49	53	37	47	87	73	68	86
	OA = 0.68	Kappa = 0.62

LD ^a: Soil classification according to SiBCS [28]. UA: User’s accuracy; PA: Producer’s accuracy; OA: overall accuracy.

and

Table A15. Pixel-by-pixel internal validation confusion matrix between the predicted map generated by the ELT2 model and the legacy map (reference), in LD2.

ELT2	Reference
LD a	RQo1	RQo5	RQg5	RQg1	GXve2	ESo	LAd	PAd	GZn1	Total	UA %
RQo1	51,998	3515	0	0	4438	2792	0	0	16,186	78,929	66
RQo5	12	56,228	0	2737	8501	672	6270	8043	35	82,498	68
RQg5	59	17	1004	0	0	0	0	0	3768	4848	21
RQg1	22	7177	0	36,231	100	0	43,762	6379	1	93,672	39
GXve2	6689	31,827	0	510	34,703	6174	56	2867	5389	88,215	39
ESo	5524	1753	0	0	9625	36,921	0	0	2675	56,498	65
LAd	3	14,609	0	30,065	4	0	175,185	19,718	1	239,585	73
PAd	0	5024	0	14,864	200	0	7121	74,914	0	102,123	73
GZn1	1278	9	410	0	4444	3070	0	0	159,276	168,487	95
Total	65,585	120,159	1414	84,407	62,015	49,629	232,394	111,921	187,331	914,855
PA %	79	47	72	43	56	74	75	67	86
	OA = 0.69	Kappa = 0.63

LD ^a: Soil classification according to SiBCS [28]. UA: User’s accuracy; PA: Producer’s accuracy; OA: overall accuracy.

and Figure 7). The average value of OA in LD1 and LD2 was 0.71 and 0.70, respectively; however, the average $\widehat{k}$ value was 0.62 in LD1 and 0.65 in LD2. These results lead us to not indicate the aggregation of MUs. In the analysis of the F1 score (Figure 7), we observed that the individual Ranger and RF models, as well as all other models trained, in general, as MUs G in LD1 and GZn1 in LD2, obtained the highest values, indicating that all models were able to predict them.

Regarding Table 1 and Table 2, we calculated the sampling density, from which we observed that the Ranger achieved the highest F1 scores for G and GZn1, even though their densities of 0.27 and 0.18 points/km², respectively, were the second lowest among the classes studied, represented by LD1 and LD2. In our work, the Ranger model was able to capture soil variability even with low sampling density, corroborating the findings of [22], who compared the potential of decision tree and vector learning quantization models for soil class prediction, finding greater accuracy for decision tree models in areas with low altitude variation.

4.2. Uncertainty and Practical Applications

The quantified uncertainty in PDSM is one of the main advantages over traditional soil surveys, in which the spatial distribution of mapping units or soil classes is a source of error. This reflects classification uncertainty and may be related, for example, to combined mapping units whose evaluation criteria are subjective, or to estimation models composed of coexisting soils or other generalized combinations of dominant soils [2,73].

In the study area, mapping units of hydromorphic soil classes cover more than 40% of the municipality of Tracuateua and present specific challenges, such as limited accessibility for the soil survey team. Incorporating uncertainty into the existing map can substantially assist future field surveys, reducing the time and resources required between planning and map production.

Locations with a higher entropy and confusion index, both close to 1 (Figure 12 and Figure 13), are associated with greater difficulty in pedogenesis modeling, requiring a larger number of samples to improve the quality of the soil map [67]. This demonstrates, in the present study, that models related to EL can improve the prediction of soil mapping units in areas where individual algorithms presented greater uncertainty.

Commonly represented by means of maps, the incorporation of uncertainty continues to be a major challenge in pedometrics [18,74]. In the joint analysis of the F1 score (Figure 7) and Figure 12 and Figure 13, in which uncertainty (entropy and IC) was compared, lower uncertainty values for MUs under hydromorphic conditions—G, E, RQg5, Eso, GXve2, RQg1, and GZn1—were obtained using models with the EL approach compared to Ranger. These results indicate that the overall accuracy of the map should be evaluated in conjunction with the uncertainty [63].

Valley bottom areas and alluvial plains, as shown in Figure 12 and Figure 13, are extremely common and vital geomorphic environments in the Amazon, concentrating significant carbon stocks in soils [75]. However, we believe that pedogenesis in these environments is intrinsically complex, being controlled by multiple interactive processes. According to [76], this complexity often reflects polygenesis, in which soils evolve through successive or overlapping pedogenetic processes that occur at different times or under different environmental conditions within the same landscape. This intrinsic complexity, combined with the lack of detailed soil maps for decision-making, represents significant challenges for PDSM in the region, since hydromorphic mapping units have subtle boundaries and high spatial variability, requiring high-resolution terrain covariates. Ref. [77] found that soils in hydromorphic lowlands (GXbd) were classified more accurately by some machine learning algorithms, highlighting the need for fine-scale terrain covariates. Therefore, the need to use high-resolution terrain covariates to map hydromorphic pedological units is supported by the recent review in [78].

In comparing MUs under hydromorphic conditions using the EL and Ranger models (Figure 12 and Figure 13), despite the low sample density for G and GZn1 (0.27 and 0.18 points/km², respectively) compared to the other classes, lower entropy and CI values were observed for the EL approach. This indicates that the joint learning approach incorporated higher predictive quality in hydromorphic environments, i.e., lower uncertainty even with a reduced number of samples, corroborating the findings of [6] in their assessment of uncertainty in digital mapping of soil classes in the southern region of the state of Minas Gerais.

Given the challenges in the Amazon biome—such as budget constraints and difficult field access, particularly in hydromorphic areas—this study has shown promise for PDSM in the region [10]. Another promising application of PDSM is predictive extrapolation, which could significantly support ongoing soil surveys with higher resolution, thus helping to fill existing gaps in soil maps throughout the Amazon [79,80,81].

5. Conclusions

Considering the polygenetic nature of tropical soils, the most important covariates (Tmean_22k, CNBL, LST, and altitude) proved adequate for modeling MUs in northeastern Pará.

Despite the scarcity of validation points for some MUs in the legacy soil map of Tracuateua, the performance metrics indicate that the EL approach provided greater predictive stability than individual machine learning models, particularly in LD2 when dealing with MUs imbalance.

The results support the hypothesis that EL can discriminate MUs under hydromorphic conditions. In addition, spatial estimates of prediction uncertainty based on entropy and CI provide useful tools for guiding field surveys and improving soil classification in difficult-to-access hydromorphic areas.

Furthermore, reducing the level of detail of MUs in older maps to simplify modeling did not improve predictive performance and may increase the implicit uncertainty associated with conventional soil maps. Furthermore, despite the slightly higher use of terrain attributes derived from DSM in terms of learning model performance compared to DTM, the greater uncertainty (Entropy and CI) observed in the latter digital model may be related to the more accurate representation of terrain elevation.

However, our approach provides a framework for identifying the most appropriate use of legacy soil class data under regional constraints. The results highlight that mapping units, which are directly associated with natural landscape patterns, may be more effectively predicted by machine learning models than taxonomic soil classes, which represent conceptual groupings defined by classification systems. Consequently, mapping units tend to exhibit spatial patterns that are more readily captured by the environmental covariates used in digital soil mapping models.