Hybrid Machine Learning–Geostatistical Modeling of Soil Quality in Saline–Sodic Soils Under Arid Agroecosystems of Northwestern Peru
Sebastian Casas‐Niño, Gerson La Rosa Lama, Sharon Mejia, Ruth Mercado, Kenyi QuispeABSTRACT
Background
Soil fertility management in mango and banana production systems in northern Peru is limited by marked spatial heterogeneity, driven by chemical imbalances, particularly salinity, alkalinity, and sodicity.
Aim
This study developed a spatial predictive framework to assess soil quality across the agricultural area of Morropón Province, Piura, by integrating a weighted soil‐quality index (SQIw), specifically designed to diagnose salinity‐, sodicity‐, and alkalinity‐related constraints, with hybrid machine learning–geostatistical modeling.
Methods
Soil sampling was carried out in homogeneous units delimited according to slope, texture, and soil color. Within each unit, subsamples were composed of the 0–30 cm layer, resulting in 1304 observations representative of mango and banana fields. The SQIw was constructed using laboratory‐derived physicochemical soil properties combined with environmental covariates, including vegetation indices extracted from Sentinel‐2 images in Google Earth Engine. A minimum data set was defined through transformation, collinearity screening, and principal component analysis. The variables were standardized, weighted according to their contribution to the explained variance, normalized using scoring functions, and linearly integrated to calculate the index, which was classified into five quality levels. Spatial prediction was performed using ordinary kriging (OK), random forest (RF), and a hybrid RF with residual kriging (RF + RK). In random cross‐validation, OK showed the lowest performance ( R 2 = 0.245; root mean square error [RMSE] = 0.155), whereas RF improved accuracy ( R 2 = 0.315; RMSE = 0.147), and RF + RK ( R 2 = 0.309; RMSE = 0.148) achieved comparable improvements.
Results
A reduced RF + RK model based on the most relevant predictors achieved the best numerical performance ( R 2 = 0.340; RMSE = 0.144), although it was not statistically different from the other RF‐based models.
Conclusion
The resulting map supports site‐specific fertility management in mango and banana production systems in northern Peru.