|
چکیده
|
Saturated soil hydraulic conductivity (Ksat) is a key component in hydrogeology and water management. This study aimed at evaluating popular tree-based machine learning algorithms (Random forest (RF), Quantile random forest (QRF), Cubist (Cu), and Decision tree regression (DTr)) to assess the spatial distribution of Ksat in a sandy agricultural field. Soil surface reflectance derived from Landsat-8 (OLI/TIRS) (several remotely sensed data including original surface reflectances of spectral bands and 22 remote sensing indices) as well as limited ground measured soil data were used as predictor covariates. Using collected disturbed and undisturbed soil samples, physicochemical properties (Ksat, porosity, organic matter, and texture contents) were determined. Based on novel supervised feature selection, covariates including Landsat-8 spectral Band2, Band4, Band6, and Band7, normalized difference pond index (NDPI), normalized difference vegetation index (NDVI), and silt content were identified as the most important factors in Ksat prediction. Variable importance analysis showed Band7 and Band6 were explaining 54% of the total Ksat variation. The model evaluation, using five statistics cretria, Taylor diagram and the Kruskal–Wallis (KW) test, demonstrates the outperformance of RF being followed by QRF, Cu, and DTr models. Accordingly, the RF method combined with an optimized sampling approach provides the most accurate digital soil Ksat map. The uncertainty analysis showed no significant different trend in Ksat predictions by applying more dense datasets confirming original-limited data (n = 28) was efficient and sufficient to derive a reliable Ksat-map. The developed approach, which can be scaled up at a large scale, seems to be useful in modelling water flow and solute transport for water and nutrient management purposes toward precision agriculture.
|