Proximal sensors fusion and machine learning algorithm combined to improve soil compaction prediction

Abstract

Predicting adverse factors in agricultural production, like excessive soil compaction, is crucial for taking preventive measures that reduce costs, drying time, environmental contamination from chemical analyses, and the need for destructive sampling methods. Therefore, our objective was to predict soil compaction by evaluating regression and classification models using Random Forest algorithms based on the integration of a wide range of proximal sensors. A total of 56 undisturbed soil samples were collected in PVC cylinders from two soil types: Anionic Acrudox (LVdf) and Typical Hapludox (LVAd), and subjected to five compaction levels (70 %, 80 %, 90 %, 100 %, 110 %) under laboratory conditions. During the experiment, 475 measurements were performed using one X-ray emission sensor, three electrical property sensors, and volumetric water content was estimated from saturation to drying. This process generated 9,025 observations across 19 sensor-derived variables. By integrating the sensors, robust and accurate regression models were developed using Random Forest algorithms to predict compaction degree, with R² = 0.93 when combining both soils and LVdf (R² = 0.79; RMSE = 7.18) and LVAd (RMSE = 6.35). Excluding water content did not significantly reduce model accuracy but altered the importance of certain variables such as Fe, Si, Ti, and Zn. The pXRF was better at predicting compaction compared to electrical sensors, achieving an R² = 0.78 for LVdf and LVAd. Classification models also performed well, reaching an overall accuracy of 0.92 (Kappa = 0.89), and Kappa values of 0.86 for LVdf and 0.74 for LVAd. Sensor fusion allowed variable analysis without disturbing soil structure, supporting potential large-scale spatial modeling for broader applications.