Predição decompactação em dois solos sob diferentes conteúdos de água com sensores proximais

Abstract

The prediction of adverse factors in agricultural production, such as excessive soil compaction, is crucial for taking preventive measures that reduce costs, drying time, environmental contamination, and sample destruction. In this study, undisturbed samples were collected from native vegetation (a transition between the Cerrado and Atlantic Forest in Lavras, Minas Gerais) from two soils (very clayey textured Latossolo Vermelho distroférrico and clayey textured Latossolo Vermelho Amarelho distrófico) in PVC cylinders with an approximate diameter of 11.2 cm and height of 6.8 cm, at a depth of approximately 10 cm below the surface layer. After being subjected to a tension of 10 kPa, the samples were compacted using a modified uniaxial consolidometer to different degrees of compaction (70%, 80%, 90%) and a Proctor test for (100% and 110%). Once saturated, measurements were taken every two days with proximal sensors, including X-ray fluorescence (with chemical elements in percentages: Mg, Al, Si, P,K, Ca, Ti, V, Cr, Mn, Fe, Ni, Cu, Zn, As, Zr), electrical resistivity (ρ) with an X5xtal 250 (in Ωm), electrical conductivity (EC) with a Teros12 (in μS cm-1), dielectric constant (Ka) with an ML2X (Ɛ dimensionless), and volumetric water content (θv in m3 m-3), until the samples dried. A Random Forest algorithm was used to: (1) predict the degree of compaction through a regression model with data fusion from proximal sensors; (2) test the effect of excluding soil water content in predicting soil compaction with proximal sensors; (3) compare the potential of soil compaction prediction models generated by sensors of electrical properties (ρ, EC, Ka) with the pXRF sensor; and (4) predict soil compaction using Random Forest classification models. The regression model with Random Forest produced robust predictions. Despite having a smaller sample size (n=4600), LVdf showed better performance than LVAd (n=4900), achieving an R2=0.79, RMSE=7.18, and MAE=4.66 based on external validation. When integrating both soils (LVdf+LVAd, n=9500), the model reached R2=0.93. Although excluding water content (θv) did not significantly impact the accuracy of the models, it altered the importance of the variables, particularly Fe in LVdf and Si, Ti, Zn, and ρ in LVAd. When considering the sensors individually, the X-ray fluorescence (pXRF) sensor was better at predicting compaction compared to the electrical sensors, achieving an R2 of 0.78 for LVdf and LVAd, finally R2=0.91 when combining both soils. Although the accuracy metrics for the classification models were good, no clear pattern of variables based on the degree of compaction could be identified. However, it can be stated that certain variables lose their importance when compared between classification and regression models. The key variables identified were Si, Al, Cu, Fe, Ti, and EC.