Design and performance evaluation of a sensor-based device for monitoring soil properties

Abstract Sensors used in precision agriculture can support rapid, non-destructive, and cost-effective soil assessment for field-based decision-making. This study developed and evaluated an affordable multi-parameter sensor fusion device ($134) for measuring soil properties under contrasting field conditions. Field experiments were conducted across four evaluated soil matrices (clay, clay loam, loam, and sand) and four depth profiles (15, 30, 45, and 60 cm), with five independent field profiles per matrix, yielding 80 depth-specific observations. To assess internal predictive performance, the dataset was partitioned at the field-profile level using a stratified 70:30 internal hold-out scheme. Fourteen field profiles (56 depth-specific observations) were used for calibration, while six separate field profiles (24 depth-specific observations) were reserved for internal validation. This procedure assessed prediction on previously unseen profiles within the investigated soil matrices and sampling zones; it did not constitute external validation. Profile-based mixed-effects analysis identified statistically significant differences associated with the evaluated soil matrices, sampling depths, and their interaction (P < 0.01) for nitrogen (N), phosphorus (P), potassium (K), and moisture content (MC). The calibration relationships showed coefficients of determination ranging from R 2 = 0.8594 to 0.9996, while the corresponding sensor–laboratory measurements showed positive Pearson correlations ranging from 0.652 to 1.000. At the 15 cm depth, the highest observed N, P, and K values occurred in clay, clay loam, and loam matrices (clay: 80.0, 29.0, and 383.3 mg kg −1 ; clay loam: 61.6, 21.0, and 300.0 mg kg −1 ; loam: 46.5, 14.5, and 136.3 mg kg −1 ), compared with the evaluated sandy profiles (19.5, 9.3, and 82.0 mg kg −1 ). Soil pH ranged from 6.4 to 7.2, and moisture content was higher at the 60 cm depth. Bland–Altman analysis indicated a mean difference of 0.130 mg kg −1 , with limits of agreement from − 0.001 to 0.261 mg kg −1 . The plot showed no clear visual trend in measurement differences across the evaluated range, supporting consistency of sensor–laboratory agreement within the investigated dataset. The proposed platform provides an affordable option for site-specific soil monitoring and rapid field screening under the investigated conditions.