ABSTRACT Soil is an important carbon repository in terrestrial ecosystems, serving a fundamental role in the intricate cycling of this elemental component. Wetlands are crucial components of the global carbon cycle, playing a significant role in carbon sequestration due to their remarkable productivity and unique sedimentary structures. Our study focuses on the wetlands east of Marajó island, recognized as the largest fluvial-marine plain in South America. In this study, we applied a methodological framework to optimize SOC content prediction in the wetlands of Marajó island using readily available environmental covariates. We collected and analyzed 81 soil samples from the most representative geoenvironments on the island at a layer of 0.00-0.20 m. Our database included vegetation indices, morphometric maps, and covariates based on distance from water bodies and archaeological sites. We tested five machine learning algorithms - Cubist, Linear Model, Random Forest, K Nearest Neighbor, and Support Vector Machine - to obtain the best prediction performance. Cubist model demonstrated the highest performance for training (R2 = 0.483) and testing (R2 = 0.505) datasets, making it the optimal choice for SOC prediction in the topsoil. The most important covariates selected by Cubist using recursive feature elimination were digital elevation model, topographic heterogeneity index, vertical distance between the summit and base of the slope, and Euclidean distance from water bodies. Geoenvironments characterized by dense alluvial rainforest with palms on Plinthosols and Gleysols, mangroves with Gleysols, and coastal muddy plains exhibited the highest SOC content in the topsoil. ecosystems component cycle structures fluvialmarine fluvial marine America 8 0.000.20 000020 0.00 0.20 0 00 20 0.00-0.2 m indices maps sites Model Forest Neighbor R2 R (R 0.483 0483 483 0.505 0505 505 datasets topsoil index slope Gleysols 000 0.000.2 00002 0.0 020 0.2 2 0.00-0. 0.48 048 48 0.50 050 50 0.000. 0000 0. 02 0.00-0 0.4 04 4 0.5 05 5 0.000 0.00-

Resumen El avance de los modelos predictivos mediante Machine Learning Algorithms (ML) asociados a datos ambientales permite mejorar los modelos de fragilidad ambiental, que son herramientas fundamentales para la toma de decisiones. Este estudio tuvo como objetivo derivar una predicción de la fragilidad ambiental mediante la prueba de ML asociado con covariables ambientales en el estado de Minas Gerais. Se utilizaron variables físico-ambientales (suelo, geología, clima, relieve) con peso de fragilidad para los atributos y cálculo de la media para obtener un modelo de Fragilidad Ambiental Potencial (PEF). Posteriormente, extrajimos los valores de PEF a una cuadrícula de 4800 puntos, que se utilizó para generar una nueva predicción de ML, llamada PEFML. Esta predicción se basó en la prueba de cinco algoritmos y un conjunto de 105 covariables ambientales. Los resultados indicaron que la predicción PEFML con mejor desempeño fue el modelo Random Forest (R2 0.59 y RMSE 0.47), indicando un predominio del bajo nivel de fragilidad ambiental. Los modelos PEF y PEFML muestran fuertes correlaciones (0,7 Pearson); sin embargo, PEFML tiene correlaciones más fuertes con otros datos ambientales. Por lo tanto, la predicción PEFML es un modelo robusto que captura información de covariables y tiene patrones espaciales coherentes. (ML decisiones Gerais físicoambientales físico suelo, suelo (suelo geología clima relieve PEF. . (PEF) Posteriormente 480 puntos 10 R2 R (R 059 0 59 0.5 0.47, 047 0.47 , 47 0.47) 0,7 07 7 (0, Pearson Pearson) embargo tanto coherentes (PEF 48 1 05 5 0. 04 0.4 4 0, (0 (

Resumo O avanço de modelos preditivos por Algoritmos de Aprendizado de Máquina (ML) associados à dados ambientais possibilita aprimoramento de modelos de fragilidade ambiental, os quais são importantes ferramentas para tomada de decisão. O objetivo desse estudo foi derivar uma predição de fragilidade ambiental, testando ML associados a covariáveis ambientais no estado de Minas Gerais. Utilizamos variáveis físico-ambientais (solo, geologia, clima, relevo) com peso de fragilidade para os atributos e cálculo da média obtendo o modelo de Fragilidade Ambiental Potencial (PEF). Posteriormente, extraímos os valores de PEF para uma grade de 4.800 pontos e usadas para gerar uma nova predição por ML, denominada PEFML. A predição foi com teste de cinco algoritmos e conjunto de 105 covariáveis ambientais. Comparamos os dois modelos de fragilidade ambiental (PEF e PEFML), inclusive com outros dados de riscos/vulnerabilidade/fragilidade. Os resultados indicaram que a predição de PEFML de melhor desempenho foi o modelo Random Forest (R2 0.59 e RMSE 0.47), indicando predomínio do nível fragilidade baixa. Os modelos de fragilidade PEF e PEFML têm forte correlação (0.7 Pearson), porém, PEFML possui correlações mais fortes com outros dados ambientais. Portanto, a predição de PEFML é um modelo robusto que capta informações de covariáveis e possui padrões espaciais coerentes (ML decisão Gerais físicoambientais físico solo, solo (solo geologia clima relevo PEF. . (PEF) Posteriormente 4800 4 800 4.80 10 PEFML, , PEFML) riscosvulnerabilidadefragilidade riscos vulnerabilidade riscos/vulnerabilidade/fragilidade R2 R (R 059 0 59 0.5 0.47, 047 0.47 47 0.47) baixa 0.7 07 7 (0. Pearson, Pearson Pearson) porém Portanto 480 80 4.8 1 05 5 0. 04 0.4 (0 48 8 4. (

Abstract The advancement of predictive models by Machine Learning Algorithms (ML) associated with environmental data enables the improvement of models of environmental fragility, which are essential tools for decision-making. This study aimed to derive a prediction of environmental fragility by testing ML associated with environmental covariates in the state of Minas Gerais. We use physical-environmental variables (soil, geology, climate, relief) with a weight of fragility for the attributes and calculation of the average to obtain a model of Potential Environmental Fragility (PEF). Subsequently, we extracted the PEF values to a 4,800-point grid, which was used to generate a new prediction by ML called PEFML. This prediction was based on testing five algorithms and a set of 105 environmental covariates. The results indicated that the best-performing PEFML prediction was the Random Forest model (R2 0.59 and RMSE 0.47), indicating a predominance of the low environmental fragility level. The PEF and PEFML models have strong correlations (0.7 Pearson); however, PEFML has stronger correlations with other environmental data. Therefore, the PEFML prediction is a robust model that captures information from covariates and has coherent spatial patterns. (ML decisionmaking. decisionmaking decision making. making decision-making Gerais physicalenvironmental physical soil, soil (soil geology climate relief PEF. . (PEF) Subsequently 4,800point 4800point point 4,800 4 800 grid 10 bestperforming best performing R2 R (R 059 0 59 0.5 0.47, 047 0.47 , 47 0.47) level 0.7 07 7 (0. Pearson Pearson) however Therefore patterns (PEF 800point 4800 4,80 80 1 05 5 0. 04 0.4 (0 480 4,8 8 ( 48 4,

ABSTRACT: The great geological and soil variation in the state of Minas Gerais, Brazil, indicates the need for regional studies to understand the geochemical background of soils. The Rio Doce Basin became a priority area for geochemical background determination after the rupture of the tailings dam of Fundão in 2015. In this context, the objectives of this study were to propose Reference Values of Soil Quality in the Rio Doce Basin, to define variables that can predict metal(loid) concentrations in the soil, and to examine the correlation between metal(loid) concentrations determined by X-ray fluorescence and by the traditional method. One hundred and seven samples were collected from minimally disturbed areas, representing the main soils and source materials. Metal(loid)s were determined by acid digestion and X-ray fluorescence. Descriptive statistics of the data, as well as the calculation of the Randomized Dependence Coefficient (RDC) and Principal Component Analysis (PCA) were carried out. The soils were found to be acidic, dystrophic with low Mehlich-1 extracted P contents, and have a variable texture. The coefficient of determination ranged from 0.4 to 0.9, suggesting X-ray fluorescence as a promising technique for determining metal(loid) concentrations in soils. The absence of correlation between clay and organic matter contents with metal(loid) concentrations suggests that the latter were inherited exclusively from the parent material, with little influence of pedogenesis. Metal mineralization in the highlands that constitute the topographic drainage divide of the basin increase the reference values of soil quality to higher values than established for the State of Minas Gerais.

ABSTRACT: Ultisols are the most common soil order in the Brazilian Amazon. The Legal Amazon (LA) has an area of 5 × 106 km2, with few accessible areas, which restricts studies of soils at a detailed level. The pedological properties can be estimated more efficiently using statistical procedures and machine learning techniques, tools which are capable of recognizing patterns in a large soil database. We analyzed the main chemical and physical properties of the B horizons of the Ultisols of the Brazilian Amazon, as well as the spatial variability of the most explanatory properties of these horizons. Physical and chemical data of 1,068 profiles of the RadamBrasil Project were used. A principal component analysis (PCA) was applied and the most explanatory variables were separated by morphostructural units and climate zones. The technique of machine learning was used for spatialization of the explanatory variables based on predictive covariates. In general, the horizons are thick, clay, with a predominance of negative charges, and low levels of exchangeable cations. The variables retained in the PCA were: sum of bases (SB), Al3+, degree of flocculation (Floc), ∆pH, and organic carbon content (C). Areas of greater precipitation have low SB, with higher values in the basement complex (BC) and in areas under the Andean influence. Higher levels of Al3+ and degrees of flocculation were also associated with greater precipitation. However, the soils are predominantly electronegative, showing a kaolinitic mineralogy. The C contents in general were low, with an increase in more humid zones due to the process of mineralization and illuviation (podzolization), and in the BC due to the protection of C by the aggregation of clay. The use of multivariate analysis allowed a better understanding of the Ultisols’ main properties in different morphostructural and climatic domains, and its spatialization facilitated the interpretation of properties and their relationships with environmental characteristics in the Legal Amazon.