Identifying hidden factors influencing soil Olsen-P in an alkaline calcareous soil using machine learning and geostatistical techniques

Abstract

Phosphorus (P) deficiency is one of the major constraints for sustainable crop production in calcareous soils. This study aimed to elucidate the key soil characteristics modulating the variability of soil Olsen P in these typical soils. A comprehensive soil sampling initiative (1.5 samples per hectare) was conducted on a 100-ha farm, considering 31 attributes that included soil physical and chemical properties, and geographic attributes. Three machine learning algorithms—partial least squares regression (PLSR), random forest (RF), and cubist regression (CR)—were employed to understand key variables controlling soil Olsen P. Furthermore, the same data set was used to spatially map the variations in Olsen P levels using ordinary kriging. The results revealed that soil chemical factors, specifically exchangeable manganese and zinc, cation exchange capacity, and carbonate, played a crucial role in controlling P levels. Among the machine learning models, the best performing model was RF (R2 = 0.95, RMSE = 1.30 mg kg−1) followed by CR (R2 = 0.92 and RMSE = 1.43 mg kg−1). Additionally, the analysis using a Gaussian semi-variogram model showed a good performance (R2 = 0.78, RMSE = 2.05 m) in visualizing the spatial distribution of Olsen P, revealing its heterogeneity. The resulting pattern of Olsen P distribution may be attributed not only to soil properties but also to external factors, such as sediment transport through watercourses across the study area and atmospheric deposition from a nearby P mining site. Overall, the combination of geostatistical methods and machine learning approach demonstrates a significant potential in understanding the complexity of soil available P (Olsen-P) that could help to develop sustainable and precise P management.

Keywords

• Kriging
• Soil mapping
• Random forest
• Cubist regression
• Partial least squares regression
• Calcareous soil