Given the critical importance of accurate soil property data in agricultural management, remote sensing offers a promising alternative to traditional field sampling an‎d laboratory analysis, which are often costly an‎d time-consuming. This study aimed to estimate key physical an‎d chemical properties of soil—namely particle size distribution, salinity, pH, organic matter content, an‎d bulk density—using satellite-derived data from Sentinel an‎d the global SoilGrids database, an‎d to compare the estimates with laboratory-measured values. In addition, the influence of these properties on the calculation of plant-available water was assessed. Soil sampling was conducted across eight regions in Iran, including Golpayegan, Ardestan, Semirom, Shahreza, an‎d Alavijeh (Isfahan province); Fariman (Razavi Khorasan); Arak (Markazi); an‎d Abadeh (Fars). A total of 189 soil samples were collected from the 0–30 cm depth range an‎d analyzed in the laboratory. Particle size distribution was determined using the hydrometer method; salinity an‎d pH were measured in saturated paste extracts; organic matter content was analyzed via the Walkley-Black method; an‎d bulk density was measured using the core method. Remote sensing data were obtained through the Google Earth Engine platform an‎d the SoilGrids global database. Statistical comparisons were carried out using paired t-tests an‎d linear regression models. Two machine learning models—Multilayer Perceptron (MLP) neural networks an‎d AdaBoost—were developed to estimate particle size distribution an‎d soil salinity, respectively. Results indicated that the accuracy of satellite data in estimating soil properties varied depending on the property type, data source, an‎d geographic location. Sentinel imagery outperformed SoilGrids in estimating particle size fractions, particularly san‎d an‎d silt, while it showed lower accuracy for clay. Conversely, SoilGrids data were more reliable for estimating pH an‎d organic matter content. Regarding available water estimation, analysis of error between two ran‎dom subsets of measured samples revealed a variation of 4–7%, leading to the adoption of a 7% error threshold as an acceptable margin for satellite-derived available water estimates. Approximately 20–23% of satellite-based estimates fell within this threshold. Applying regression models to harmonize measured an‎d remotely sensed data significantly improved available water estimation accuracy in most regions. Specifically, the error margin decreased by 61% using Sentinel data an‎d by 67% with SoilGrids data. The machine learning model also demonstrated strong performance in predicting soil texture: for clay, the model yielded an R² of 0.76 an‎d an RMSE of 4.27%; for san‎d, an R² of 0.73; an‎d for silt, a slightly lower but still acceptable R² of 0.65. The AdaBoost model for predicting soil salinity achieved an R² of 0.65 across all study areas, with considerably higher accuracy in the Isfahan region, where R² reached 0.97 an‎d RMSE was as low as 1.35 dS/m, confirming the model’s robustness.

مهدي قيصري

محمدرضا مصدقي , محبوبه قبادي

محمد شايان نژاد , حسين شريعتمداري