Mapping topsoil pH using different predictive models and covariate sets in Henan Province, Central China

Abstract

Spatial prediction and mapping of soil properties are synergistically influenced by models and environmental covariates. However, selecting the best modeling technique for a given landscape has been a challenge for mapping soil property. Typically, the conventional variable set (such as soil factors, climatic factors, terrain factors, and organism factors) is widely used to predict soil pH. However, certain covariates are difficult to obtain for some regions with poor data. Additionally, the effectiveness of vegetation indices (such as normalized difference vegetation index, NDVI) and phenological variables that can monitor vegetation growth dynamics based on satellite remote sensing for soil pH mapping are largely unclear. Accordingly, this study aimed to set up five environmental variable sets, such as NDVI, phenological variables, conventional variable set, conventional variable set + NDVI (C + N), and conventional variable set + phenological variables (C + P). The random forest (RF) model and convolutional neural network (CNN) model were used to compare the performance of these five environmental variable sets for spatial prediction for topsoil pH. In total, 4332 soil samples were applied and randomly divided into calibration (80%) and validation sets (20%) in this study. The model performance was evaluated using the coefficient of determination (R2) and root mean square error (RMSE). The results showed that among the five environmental variable sets, the conventional variable set was the best to predict soil pH (R2 = 0.85, RMSE = 0.41). Further, among the remote sensing variables, the phenological variables outperformed NDVI, explaining up to 67% of the variation in topsoil pH. The RF model could achieved higher accuracy than the CNN model regardless of the variables selected. Moreover, the conventional variable set could predict topsoil pH, and exhibited the potential of phenological variables. The results indicated when the conventional variable set was difficult to obtain or was missing, phenological variables can be used to predict soil properties and perform digital mapping instead of the conventional variable set.