Texture prediction of natural soils in the Brazilian Amazon through proximal sensors

Proximal sensors provide fast, low-cost, environmentally friendly, and reliable analyses for the characterization of soils and other materials. Numerous studies have been conducted on soils in temperate regions, but there are knowledge gaps regarding the use of these devices in tropical soils, especially in the Amazon region. In this regard, this study utilized portable proximal sensors of X-ray fluorescence spectroscopy (pXRF) and diffuse reflectance spectroscopy in the visible to near-infrared region (Vis-NIR) for predicting the texture of natural soils in 61 municipalities in the state of Pará, Amazon region, Brazil. The objectives were: i) to investigate the accuracy of soil texture prediction based on data from sensors separately and sensor fusion (pXRF and Vis-NIR data) using two supervised algorithms (Random Forest, RF, and Support Vector Machine, SVM) and ii) to assess the effect of soil horizon (superficial and subsuperficial horizons, and their combination) in predicting the texture of tropical natural soils. In total, 233 soil samples were collected in the 0–20 cm and 80–100 cm depths, equivalent to superficial and subsuperficial horizons in areas with primary or secondary forest cover with at least 20 years of natural regeneration and approximately 20 ha of coverage area. The hydrometer method was used for soil texture analysis. In parallel, a portion of the soil samples was analyzed by pXRF and Vis-NIR, in triplicate, under laboratory conditions. The predictive models with RF were more robust compared to the models obtained with SVM, according to ratio performance interquartile distance (RPIQ), coefficient of determination (R2), root mean square error (RMSE), and mean absolute error (MAE). The R2 values obtained by pXRF, Vis-NIR, and sensor data combination were, respectively, 0.89, 0.87, and 0.93 for sand; 0.92, 0.90, and 0.93 for clay; and 0.91, 0.67, and 0.93 for silt. Overall, clay prediction models achieved higher R2 values compared to sand and silt models. Soil texture prediction using sensor fusion showed lower RMSE values and higher R2 and RPIQ values, respectively (sand: 7.79, 0.93, 4.69; clay: 5.58, 0.93, 3.86; and silt: 5.72, 0.92, 2.92) compared to the best-performing sensor individually (Vis-NIR). With regard to the optimal model utilizing individual sensor data, Vis-NIR models exhibited reduced error for clay and sand prediction. The effect of combining horizons to a single and bigger dataset was minimally important for the model