Predição da adsorção de fósforo e matéria orgânica em solos cultivados de Minas Gerais utilizando sensores proximais e atributos de fertilidade do solo

Abstract

In highly weathered tropical soils, phosphorus adsorption and low available P content, as well as organic matter (SOM) dynamics, are important factors for fertility management. Soil attributes used to assess P and SOM typically require expensive, labor-intensive, and environmentally polluting laboratory methods. Currently, many soil attributes can be evaluated using proximal sensors, offering advantages such as speed and the absence of chemical reagents. Thus, the objective of this study was to combine portable X-ray fluorescence (pXRF) data with soil color parameters obtained through smartphones and portable digital sensors to develop predictive models for indicators of P adsorption and availability, as well as SOM content in soils. A total of 108 surface samples (0-20 cm) from agricultural areas in Minas Gerais were utilized in this study. These samples were employed for elemental analysis via pXRF, conducting P adsorption tests, determining residual phosphorus (P-rem), available P, and organic matter content (SOM). To obtain color data (RGB) from the samples, two smartphones with different operating systems and under varying lighting conditions were employed, alongside the Nix Pro proximal sensor. The obtained results underwent descriptive statistical analysis, principal component analysis (PCA), and Pearson correlations to assess the influence of chemical, physical, mineralogical, and organic matter composition on parameters related to P adsorption. Moreover, Pearson correlation was utilized to evaluate the performance of the different devices in obtaining RGB parameters. The Random Forest algorithm was employed to create prediction models. The best-selected models exhibited the lowest root mean square error (RMSE) values and higher residual prediction deviation (RPD) and coefficient of determination (R 2 ) values. The results demonstrated that the combination of data obtained via pXRF + color data allowed for the development of prediction models with satisfactory accuracy for CMAP (RPD between 1.60 and 1.74), available P (RPD between 2.36 and 2.52), and SOM (RPD between 2.54 and 2.78). Fertility data combined with color yielded accurate prediction models for SOM (RPD between 2.61 and 2.87) and moderately accurate models for P-rem (RPD between 1.67 and 1.92) and available P (RPD between 1.41 and 1.49). The pXRF + soil fertility model enabled the best predictions for SOM (RMSE = 0.36; R 2 = 0.90 and RPD = 3.26) and P-rem (RMSE = 5.14; R 2 = 0.82 and RPD = 2.19). For CMAP, the pXRF + colors obtained by the iPhone in natural shade yielded the best coefficients (RMSE = 284.54; R 2 = 0.71 and RPD = 1.74), while for available P, the model using only pXRF data was the most accurate (RMSE = 13.83; R 2 = 0.85 and RPD = 2.53). Reliable models for the K L index were not obtained. Therefore, the use of proximal sensors and digital images showed promising potential for predicting parameters indicative of P adsorption and availability, as well as SOM.