ABSTRACT Multitemporal collections of satellite images and their products have recently been explored in digital soil mapping. This study aimed to produce a bare soil image (BSI) for the São Paulo State (Brazil) to perform a pedometric analysis for different geographical levels. First, we assessed the potential of the BSI for predicting the surface (0.00-0.20 m) and subsurface (0.80-1.00 m) clay, iron oxides (Fe 2 O 3 ), aluminum (m%) and bases saturation (V%) contents at the state level, which are important properties for soil classification. In this task, legacy soil samples, the BSI and terrain attributes were employed in machine learning. In a second moment, we evaluated the capacity of the BSI for clustering the landscape at the regional level, comparing the predicted patterns with a legacy semi-detailed soil map from a smaller reference site. In the final stage, the predicted soil maps from the state level were investigated at the farm level considering several sites distributed across the São Paulo state. Our results demonstrated that clay and Fe 2 O 3 reached the best prediction performance for both depths at the state level, reaching a RMSE of less than 10 %, RPIQ higher than 1.6 and R 2 of at least 0.41. Additionally, the predicted landscape clusters had a significant association with the main pedological classes, subsurface color, soil mineralogy and texture from the legacy semi-detailed soil map. Illustrative examples at the farm level indicated great capacity of BSI in detecting the variations of soils, which were linked to several soil properties, such as texture, iron content, drainage network, among others. Therefore, this study demonstrates that BSI is valuable information derived from optical Earth Observation data that can contribute to the future of soil survey and mapping in Brazil (PronaSolos).

Resumo: O objetivo deste trabalho foi desagregar os polígonos de mapas de unidades fisiográficas, de modo a individualizar as classes de solos ocorrentes em cada unidade, para representá-las como unidades de mapeamento simples de solos e gerar um mapa de solos com maior detalhe cartográfico que o mapa original, ampliando a utilidade desses dados em demandas futuras. Foi utilizado um mapa fisiográfico, em escala 1:25.000, da microbacia Córrego Tarumãzinho, localizada no Município de Águas Frias, no Estado de Santa Catarina. Para realizar a desagregação, foram utilizados três parâmetros geomorfométricos: declividade e formas do terreno, ambas derivadas do modelo digital do terreno; e mapa de elevação. Os limites das unidades fisiográficas e os mapas de elevação, declividade e formas do terreno foram submetidos à tabulação cruzada para identificar as combinações existentes entre as classes de solos que compõem cada unidade fisiográfica. A partir dessas combinações, foram elaboradas regras para selecionar áreas de ocorrência típica de cada tipo de solo, para treinar um modelo de árvores de decisão para predição da ocorrência das classes de solos. O treinamento do modelo foi realizado no programa Weka, e a sua validação foi feita com um conjunto de perfis de solos georreferenciados. A desagregação possibilita a individualização e a espacialização das classes de solos e é útil para a produção de mapas de solos detalhados.

Abstract: The objective of this work was to disaggregate the polygons of physiographic map units in order to individualize the soil classes in each one, representing them as simple soil map units and generating a more detailed soil map than the original one, making these data more useful for future reference. A physiographic map, on a 1:25,000 scale, of the Tarumãzinho watershed, located in the municipality of Águas Frias, in the state of Santa Catarina, Brazil, was used. For disaggregation, three geomorphometric parameters were applied: slope and landforms, both derived from the digital terrain model; and an elevation map. The boundaries of the physiographic units and the elevation, slope, and landform maps were subjected to cross tabulation to identify the existing combinations between the soil classes of each physiographic unit. Based on these combinations, rules were established to select typical areas of occurrence of each soil type in order to train a decision tree model to predict the occurrence of soil classes. The model was trained using the Weka software and was validated with a set of georeferenced soil profiles. Disaggregation enables the individualization and spatialization of soil classes and is useful in producing detailed soil maps.

ABSTRACT Soil surveys often contain multi-component map units comprising two or more soil classes, whose spatial distribution within the map unit is not represented. Digital Soil Mapping tools supported by information from soil surveys make it possible to predict where these classes are located. The aim of this study was to develop a methodology to increase the detail of conventional soil maps by means of spatial disaggregation of multi-component map units and to predict the spatial location of the derived soil classes. Three digital maps of terrain variables - slope, landforms, and topographic wetness index - were correlated with the soil map and 72 georeferenced profiles from the Porto Alegre soil survey. Explicit rules that expressed regional soil-landscape relationships were formulated based on the resulting combinations. These rules were used to select typical areas of occurrence of each soil class and to train a decision tree model to predict the occurrence of individualized soil classes. Validation of the soil map predictions was conducted by comparison with available soil profiles. The soil map produced showed high agreement (80.5 % accuracy) with the soil classes observed in the soil profiles; Ultisols and Lithic Udorthents were predicted with greater accuracy. The soil variables selected in this study were suitable to represent the soil-landscape relationships, suggesting potential use in future studies. This approach developed a more detailed soil map relevant to current demands for soil information and has potential to be replicated in other areas in which data availability is similar.

ABSTRACT A large number of predictor variables can be used in digital soil mapping; however, the presence of irrelevant covariables may compromise the prediction of soil types. Thus, algorithms can be applied to select the most relevant predictors. This study aimed to compare three covariable selection systems (two filter algorithms and one wrapper algorithm) and assess their impacts on the predictive model. The study area was the Lajeado River Watershed in the state of Rio Grande do Sul, Brazil. We used forty predictor covariables, derived from a digital elevation model with 30 m resolution, in which the three selection models were applied and separated into subsets. These subsets were used to assess performance by applying four prediction algorithms. The wrapper method obtained the best performance values for the predictive model in all the algorithms evaluated. The three selection methods applied reduced the number of covariables in the predictive models by 70 % and enabled prediction of the 14 soil mapping units.

Uma questão crítica no mapeamento digital de solos é a seleção do método de amostragem dos dados para treinamento do modelo preditivo. Uma abordagem emergente aplica a seleção de instâncias (observações) para reduzir o tamanho do conjunto de dados, selecionando amostras relevantes para obter um subconjunto representativo, o qual seja grande o suficiente para preservar as informações pertinentes, mas pequeno o suficiente para ser facilmente manipulado pelos algoritmos de aprendizagem. Embora existam sugestões para distribuir a amostragem de dados em função da proximidade de limites de unidades de mapeamento de solos (UM), ainda existem contradições entre as recomendações de pesquisa para localizar amostras mais perto ou mais distantes desses limites. Foi realizado um estudo para avaliar os métodos de seleção de instâncias com base na coleta de dados espacialmente explícita usando a localização em relação aos limites de mapa de solo como o principal critério. Realizou-se análise de árvore de decisão para a modelagem de mapeamento digital de classes de solo usando dois esquemas de amostragem diferentes: a) selecionando pontos de amostragem localizados fora das áreas marginais aos limites das UM e b) selecionando pontos de amostragem situados dentro das áreas marginais aos limites das UM. Os dados foram preparados para a geração de árvores de classificação para incluir somente dados pontuais localizados dentro ou fora de faixas com larguras de 60, 120, 240, 360, 480 e 600m ao redor dos limites de UM. Ambos os métodos de seleção de instâncias foram eficazes para reduzir o tamanho do conjunto de dados usado para calibração de árvores de classificação, mas não trouxeram vantagens para o mapeamento digital de classes de solos.

A critical issue in digital soil mapping (DSM) is the selection of data sampling method for model training. One emerging approach applies instance selection to reduce the size of the dataset by drawing only relevant samples in order to obtain a representative subset that is still large enough to preserve relevant information, but small enough to be easily handled by learning algorithms. Although there are suggestions to distribute data sampling as a function of the soil map unit (MU) boundaries location, there are still contradictions among research recommendations for locating samples either closer or more distant from soil MU boundaries. A study was conducted to evaluate instance selection methods based on spatially-explicit data collection using location in relation to soil MU boundaries as the main criterion. Decision tree analysis was performed for modeling digital soil class mapping using two different sampling schemes: a) selecting sampling points located outside buffers near soil MU boundaries, and b) selecting sampling points located within buffers near soil MU boundaries. Data was prepared for generating classification trees to include only data points located within or outside buffers with widths of 60, 120, 240, 360, 480, and 600m near MU boundaries. Instance selection methods using both spatial selection of methods was effective for reduced size of the dataset used for calibrating classification tree models, but failed to provide advantages to digital soil mapping because of potential reduction in the accuracy of classification tree models.