Digital mapping to extrapolate the selected soil fertility attributes in calcareous soils of a semiarid region in Iran.

Purpose: Spatial variability of soil properties is considered as one of the most important reasons for the variability of crop productions. The current research was conducted to determine the capability of machine learning models for the generalization of the modeling results from the reference area, i.e., Marvdasht plain, for estimating soil fertility attributes with the aims of extrapolating the modeling results to receptor area, i.e., Bandamir, in Iran with the aid of Homosoil concepts. Materials and methods: The field study consists of 200 and 50 soil sampling locations in the reference and the receptor areas, respectively. Then, four soil properties, namely, soil organic carbon, total nitrogen, available phosphorus, and exchangeable potassium were measured as soil fertility attributes. Eighty-two soil and environmental covariates were gathered from different sources, i.e., soil variables from field work and laboratory analysis, digital elevation model, and remote sensing data as potentially connected to "scorpan" factors in both areas. Here, seven soil variables (bulk density, clay, silt and sand contents, calcium carbonate equivalent, pH, and soil electrical conductivity), 36 covariates as proxy of surface and vegetation cover, and 39 attributes related to topography are representations of "s," "o," and "r" in scorpan model, respectively. Selection of the most relevant covariates, feature selection was made based on recursive feature elimination method and the performance of three ML models, i.e., cubist, random forest, and k-nearest neighbors, in the reference area for finding the best spatial model to apply in the receptor area. Results and discussion: The feature selection results showed that 15 soil and environmental covariates were most relevant and therefore were chosen as predictors of soil fertility attributes. The validation results showed that the soil fertility attribute maps in the receptor area were highly accurate than the prediction maps of the reference area based on the coefficient of determination for soil organic carbon in topsoil (R² = 0.74) and subsoil (R² = 0.52), exchangeable potassium in topsoil (R² = 0.77) and subsoil (R² = 0.27), and total nitrogen at the topsoil (R² = 0.28) and subsoil (R² = 0.25), also for available phosphorus, the prediction accuracy increased in the receptor area. The machine learning performances also showed that cubist outperformed the model in the prediction of soil fertility attributes in both areas. Also, the results revealed that soil variables identified as high-rank in the prediction of soil fertility attributes followed by topographic attributes. Conclusion: Generally, in areas where soil data is limited and detailed maps of soil properties are not available, extrapolation approach could be applied as an easy and quick methodology for preparing soil maps with low-cost and acceptable accuracy. Also, we are suggesting that the applied methodology employed in this study can be applied in another region with similar conditions. [ABSTRACT FROM AUTHOR]