Integrating machine learning with spatial analysis for enhanced soil interpolation: Balancing accuracy and visualization

Soil interpolation at the village scale plays a critical role in precision agriculture, enabling accurate predictions of soil properties across unsampled locations. Traditional interpolation methods, such as Inverse Distance Weighting (IDW), Radial Basis Function (RBF), Ordinary Kriging (Krig), and Empirical Bayesian Kriging (EBK), often face challenges in addressing the spatial heterogeneity and variability of soil data due to their linearity and stationarity assumptions and difficulties in handling nonlinear relationships. These limitations highlight the need to optimize soil interpolation approaches for improved accuracy, particularly at smaller scales. This study investigated the potential of integrating machine learning techniques with traditional interpolation methods to enhance prediction performance. The research collected data from two village scale study areas, comprising 208 and 309 sampling points, respectively. For each soil variable, traditional interpolation methods were compared to machine learning-enhanced approaches, including Multifactor Kriging (MFK), Random Forest Kriging (RFOK), and Random Forest Inverse Distance Weighting (RFIDW). The performance of these methods was evaluated based on prediction efficiency and consistency analysis. Results demonstrated that the accuracy of interpolation methods was significantly influenced by the spatial correlation and variability of soil data. Overall, machine learning-integrated methods improved prediction efficiency by over 60%. However, traditional Kriging methods produced smoother visualizations on interpolated maps compared to RFOK, which may be advantageous for specific applications, especially for large-scale applications. Consistency analysis revealed moderate similarity between traditional Kriging and IDW (Pearson correlation coefficients of 0.83 to 0.93), whereas RFOK and RFIDW exhibited stronger similarity (Pearson correlation coefficients of 0.96 to 0.98). This study underscores the innovative potential of integrating machine learning into soil interpolation, offering a significant contribution to advancing exploration of soil spatial heterogeneity. Through balancing the data accuracy and visualization effects, the findings provide actionable insights for optimizing soil management strategies and enhancing spatial prediction capabilities in agricultural research.