RESUMO: Dentre os constituintes do solo, especial atenção é voltada aos teores de argila e de matéria orgânica do solo (MOS), pois, entre outros aspectos, são determinantes para retenção de nutrientes e a formação de agregados no solo, os quais afetam diretamente o potencial produtivo das culturas. Os métodos mais comumente utilizados para quantificação destes constituintes apresentam algumas desvantagens, como o uso de reagentes químicos e a geração de resíduos. Uma alternativa a estes métodos é o uso da espectroscopia no infravermelho próximo (near infrared spectroscopy - NIRS). O objetivo deste trabalho é avaliar modelos de quantificação dos teores de argila e de MOS em amostras de solo utilizando dados espectrais por meio da técnica NIRS. Foram utilizadas 400 amostras de solos com amplitude nos teores de MOS e argila para geração de uma curva de calibração. A argila foi determinada pelo método do densímetro e a MOS por meio da solução sulfocrômica. Para calibração, utilizou-se os espectros originais (absorbância) e com pré-tratamento espectral (Savitzky-Golay derivative) das 400 amostras nos seguintes modelos: multiple linear regression (MLR), partial last squares regression (PLSR), support vector machine (SVM) e Gaussian process regression (GPR). A validação da curva foi realizada com o modelo que apresentou melhor desempenho na calibração (SVM) de duas maneiras: com 40 amostras aleatórias oriundas daquelas utilizadas na calibração e com outras 200 novas amostras desconhecidas. O teor de argila das amostras de solo afeta a capacidade preditiva da curva de calibração da estimativa do teor de MOS pelo NIRS. A validação das curvas apresentou pior desempenho (menor R² e maior RMSE) quando feita a partir de amostras desconhecidas, cujo modelo tende a superestimar os teores mais baixos e subestimar os teores mais elevados de argila e MOS com a curva gerada. Apesar do potencial de predição destes atributos via NIRS, outros estudos de calibração ainda são necessários para que esta técnica possa ser utilizada em laboratórios de análises de solos.

ABSTRACT: Among the soil constituents, special attention is given to soil organic matter (SOM) and clay contents, since, among other aspects, they are key factors to nutrient retention and soil aggregates formation, which directly affect the crop production potential. The methods commonly used for the quantification of these constituents have some disadvantages, such as the use of chemical reactants and waste generation. An alternative to these methods is the near-infrared spectroscopy (NIRS) technique. The aim of this research is to evaluate models for SOM and clay quantification in soil samples using spectral data by NIRS. A set (n = 400) of soil samples previously analyzed by traditional methods were used to generate a NIRS calibration curve. The clay content was determined by the hydrometer method while SOM content was determined by sulfochromic solution. For calibration, we used the original spectra (absorbance) and spectral pretreatment (Savitzky-Golay smoothing derivative) in the following models: multiple linear regression (MLR), partial last squares regression (PLSR), support vector machine (SVM) and Gaussian process regression (GPR). The curve validation was performed with the SVM model (best performance in the calibration based on R² and RMSE) in two ways: with 40 random samples from the calibration set and another set with 200 new unknown samples. The soil clay content affects the predictive ability of the calibration curve to estimate SOM content by NIRS. Validation curves showed poorer performance (lower R² and higher RMSE) when generated from unknown samples, where the model tends to overestimate the lower levels and to underestimate the higher levels of clay and SOM. Despite the potential of NIRS technique to predict these attributes, further calibration studies are still needed to use this technique in soil analysis laboratories.

Resumo: O objetivo deste trabalho foi avaliar o uso da seleção de covariáveis por conhecimento especializado no desempenho de modelos de predição de classes de solos em uma paisagem complexa, para identificar o melhor modelo preditivo para o mapeamento digital de solos na região Sul do Brasil. Um total de 164 pontos foram amostrados em campo, com uso do hipercubo latino condicionado, tendo-se considerado as covariáveis elevação, declividade e aspecto. A partir do modelo digital de elevação, extraíram-se as covariáveis ambientais que compuseram três conjuntos, formados por: 21 covariáveis, covariáveis após exclusão das multicolineares e covariáveis escolhidas por conhecimento especializado. A predição foi realizada com os seguintes modelos: árvore de decisão, floresta aleatória, regressão logística múltipla e máquina de vetor de suporte. A acurácia dos modelos foi avaliada pelo índice kappa (K), pela acurácia geral (AG) e pela acurácia da classe. Os modelos de previsão foram sensíveis à amostragem desproporcional de classes de solo. O melhor mapa predito obteve AG de 71% e K de 0,59. O uso do conjunto de covariáveis escolhido pelo conhecimento especializado melhora o desempenho do modelo em prever as classes de solo em uma paisagem complexa, e floresta aleatória é o melhor modelo para previsão espacial das classes de solo.

Abstract: The objective of this work was to evaluate the use of covariate selection by expert knowledge on the performance of soil class predictive models in a complex landscape, in order to identify the best predictive model for digital soil mapping in the Southern region of Brazil. A total of 164 points were sampled in the field using the conditioned Latin hypercube, considering the covariates elevation, slope, and aspect. From the digital elevation model, environmental covariates were extracted, composing three sets, made up of: 21 covariates, covariates after the exclusion of the multicollinear ones, and covariates chosen by expert knowledge. Prediction was performed with the following models: decision tree, random forest, multiple logistic regression, and support vector machine. The accuracy of the models was evaluated by the kappa index (K), general accuracy (GA), and class accuracy. The prediction models were sensitive to the disproportionate sampling of soil classes. The best predicted map achieved a GA of 71% and K of 0.59. The use of the covariate set chosen by expert knowledge improves model performance in predicting soil classes in a complex landscape, and random forest is the best model for the spatial prediction of soil classes.

ABSTRACT Soil reflectance spectroscopy has become an innovative method for soil property quantification supplying data for studies in soil fertility, soil classification, digital soil mapping, while reducing laboratory time and applying a clean technology. This paper describes the implementation of a Graphical User Interface (GUI) using R named AlradSpectra. It contains several tools to process spectroscopic data and generate models to predict soil properties. The GUI was developed to accomplish tasks such as perform a large range of spectral preprocessing techniques, implement several multivariate calibration methods, generate statistics assessment and graphical output, validate the models using independent dataset, and predict unknown variables using soil spectral data. AlradSpectra has four main modules: Import Data, Spectral Preprocessing, Modeling, and Prediction. The implementation of AlradSpectra is demonstrated by applying visible near-infrared reflectance spectroscopy for soil organic carbon (SOC) prediction. The data contains the value of SOC and Vis-NIR reflectance for 595 soil samples. The prediction statistic assessment of SOC was performed applying all spectral preprocessing and methods. The R 2 considering all models ranged from 0.54 to 0.80. In the partial least squares regression (PLSR) models, the performances were similar to multiple linear regression (MLR) and support vector machines (SVM). The lowest error in the SOC prediction was achieved by PLSR method with standard normal variate (SNV) preprocessing reaching an R 2 of 0.80, the smallest root mean square error (RMSE) of 0.47 %, and ratio of performance to inter-quartile distance (RPIQ) of 3.12. The capacity of performing multiple tasks, being free and open-source, easy to operate, and requiring no initial knowledge of R programming language are features that make AlradSpectra a useful tool to perform different modeling approaches and predict the desired soil variable.

ABSTRACT The success of soil prediction by VIS-NIR-SWIR spectroscopy has led to considerable investment in large soil spectral libraries. The aims of this study were 1) to develop a soil VIS-NIR-SWIR spectroscopy approach using legacy soil samples to improve spectral soil information in a regional scale; (2) to compare six spectral preprocessing techniques; and (3) to compare the performance of linear and non-linear multivariate models for prediction of sand and clay content. A total of 1,534 legacy soil samples, stored by Epagri, were collected from agricultural areas in 2009 on a regional scale, covering 260 municipalities of Santa Catarina. Six spectral preprocessing techniques were applied and compared with reflectance spectra (control treatment) in the development of sand and clay prediction models. Five multivariate regression models, Support Vector Machines, Gaussian Process Regression, Cubist, Random Forest, and Partial Least Square Regression were compared. The scatter-corrective preprocessing groups produced similar or better performance than spectral-derivatives. In addition, preprocessing spectra prior to regression analysis does not improve sand prediction, since reflectance spectra achieved the best performance using Cubist, SVM, and PLS models. In general, clay content presented better prediction accuracy than sand content. The best multivariate model to predict sand and clay content from soil VIS-NIR-SWIR spectra was Cubist. The best Cubist performance was achieved combined with reflectance spectra (R2 = 0.73; root mean square error = 10.60 %; ratio of the performance to the interquartile range = 2.36) and MSC (R2 = 0.83; root mean square error = 7.29 %; ratio of the performance to the interquartile range = 3.70) for sand and clay content, respectively. Considering the mean RMSE values of the validation set, the predictive ability of the multivariate models decreased in the following order: Cubist>PLS>RF>GPR>SVM for both properties. The predictive ability of VIS-NIR-SWIR reflectance spectroscopy achieved in this study for sand and clay content using legacy soil data and heterogeneous samples confirmed the potential of the spectroscopy approach.

ABSTRACT: The study of soils, including their physical and chemical properties, is essential for agricultural management. Soil quality must be maintained to ensure sustainable production of food and conservation of natural resources. In this context, soil mapping is important to provide spatial information, which can be performed using remote sensing (RS) techniques. Modeling through use of satellite data is uncertain regarding the amplitude of replicability of the models. The aim of this study was to develop a quantification model for soil texture based on reflectance information from a continuum of bare soils, obtained by overlapping multi-temporal satellite images, and apply this model to an unknown region to evaluate its applicability. Spectral data were extracted from two Landsat TM 7 satellite images containing only bare soil, representing two distinct regions in Brazil (Area 1 and Area 2). The spectral data (obtained from six bands) and laboratory data (particle size from the 0.00-0.20 m layer) of Area 1 were modeled and extrapolated to Area 2. The bare soil images differentiated textural classes as sandy, sandy loam, clayey loam, clayey, and very clayey soil. The coefficients of determination between the determined and estimated values were higher than 0.5 and errors lower than 13 % for Area 1 and 30 % for Area 2, indicating applicability of the model to unknown areas.

ABSTRACT Terrain attributes (TAs) derived from digital elevation models (DEMs) are frequently used in digital soil mapping (DSM) as auxiliary covariates in the construction of prediction models. The DEMs and information extracted from it may be limited with regard to the spatial resolution and error magnitude, and can differ in the behavior of terrain features. The objective of this study was to evaluate the quality and limitations of free DEM data and to evaluate a topographic survey (TS) underlying the choice of a more appropriate model, for use in DSMs at a scale of 1:10,000. The study was conducted in an area of 937 ha in the watershed of Lajeado Giruá, in southern Brazil. The DEMs: DEM-TS, DEM-Topographic Map (TM), DEM-ASTER, DEM-SRTM, and DEM-TOPODATA were evaluated with regard to the precision elevation by statistical tests based on field reference points, the root mean square error (RMSE), identification of the number and size of spurious depressions, and the application of the Brazilian Cartographic Accuracy Standards Law (BCASL) to define the scale of each DEM. In addition, the TA derived from each DEM was compared with the TA from DEM-TS, considered to be terrain reality. The results showed that the elevation data of DEM-TS had the best quality (RMSE = 1.93 m), followed by DEM-SRTM (RMSE = 5.95 m), DEM-Topographic Map (RMSE = 8.28 m), DEM-TOPODATA (RMSE = 9.78 m) and DEM-ASTER (RMSE = 15.57 m). The DEM-TS was well-represented at a 1:10,000 scale, while the DEM-Topographic Map and DEM-SRTM fitted 1:50,000, the DEM-TOPODATA 1:50,000 and the DEM-ASTER a 1:100,000 scale. The results of DEM-SRTM and DEM-TOPODATA were closest to terrain reality (DEM-TS) and had the lowest number of spurious depressions and RMSE values for each evaluated attribute, but were inadequate for not fitting detailed scales compatible with small areas. The techniques for the acquisition of elevation data of each DEM and mainly the flat to gently undulating topography were factors that influenced the results. For a DSM at a scale of 1:10,000 in similar areas, the most appropriate model is DEM-TS.

ABSTRACT Diffuse reflectance spectroscopy (DRS) is a fast and cheap alternative for soil clay, but needs further investigation to assess the scope of application. The purpose of the study was to develop a linear regression model to predict clay content from DRS data, to classify the soils into three textural classes, similar to those defined by a regulation of the Brazilian Ministry of Agriculture, Livestock and Food Supply. The DRS data of 412 soil samples, from the 0.0-0.5 m layer, from different locations in the state of Rio Grande do Sul, Brazil, were measured at wavelengths of 350 to 2,500 nm in the laboratory. The fitting of the linear regression model developed to predict soil clay content from the DRS data was based on a R2 value of 0.74 and 0.75, with a RMSE of 7.82 and 8.51 % for the calibration and validation sets, respectively. Soil texture classification had an overall accuracy of 79.0 % (calibration) and 80.9 % (validation). The heterogeneity of soil samples affected the performance of the prediction models. Future studies should consider a previous classification of soil samples in different groups by soil type, parent material and/or sampling region.

A espectroscopia de reflectância difusa (ERD) pode ser utilizada como alternativa para quantificação de atributos como granulometria e matéria orgânica do solo (MOS). Essa técnica pode ser opção para quantificar esses atributos em grande volume de amostras de solos, visto ser rápida, com menor custo e sem a geração de resíduos químicos. O objetivo deste estudo foi desenvolver modelos usando análise de regressão linear múltipla para predizer o teor de argila, areia, silte e MOS, utilizando dados de ERD em uma área de relevo e geologia complexa localizada na região central do Rio Grande do Sul. No estudo, foram utilizadas 303 amostras coletadas na profundidade de 0,00-0,20 m para determinar os teores de argila, areia, silte e MOS por meio da análise laboratorial e de reflectância espectral. O desempenho dos modelos de predição apresentaram bons resultados, com capacidade de explicação da variância de 77 e 72 % para areia e argila, respectivamente. Mesmo com a complexidade geológica e pedológica, os resultados evidenciaram que a técnica é promissora, sendo possível a aplicação dessa técnica para predição da granulometria e teor de MOS.

Diffuse reflectance spectroscopy (DRS) can be used as an alternative in identifying and quantifying some soil properties such as particle size and soil organic matter (SOM). This technique may be an alternative to quantifying those properties in a large volume of soil samples since it is faster and less costly and does not produce chemical residues. The aim of this study was to develop models using multiple linear regression analysis to predict the content of clay, sand, silt, and SOM using DRS data in an area of complex topography and geology located in the central region of Rio Grande do Sul, Brazil. In the study, 303 samples were collected at a depth of 0.00-0.20 m for determination of clay, sand, silt, and SOM by laboratory and spectral reflectance analysis. The predictive models produced high-quality results, explaining 77 and 72 % of the variance for sand and clay, respectively. The interpolation maps of the observed and predicted properties revealed that the spatial patterns are mainly associated with the topography and geology of the area. Even with the geological and pedological complexity, the results indicated that the technique is promising and it may be applied for prediction of particle size and SOM.