Geospatial Soil Texture Prediction in Cebu, Philippines: A Comparative Study of UK, MLR, and ANN

Abstract

Many soil maps in the Philippines are outdated, lacking the detail required for effective land management and agricultural planning. Traditional survey methods, still widely used, often fail to capture the spatial variability of soil properties, which is critical given the country’s diverse topography and land-use patterns. This study compares three models—Multiple Linear Regression (MLR), Artificial Neural Networks (ANN), and Universal Kriging (UK) for soil texture classification, using open-source software tools: R and SAGA GIS. Results indicate that the UK model significantly outperforms both MLR and ANN, with the highest overall accuracy of 77.25%, a strong kappa value (0.5757), and a statistically significant p-value (5.937e-10). The UK model excels in class-wise performance, showing superior sensitivity for key classes such as Class 1 (0.9576) and Class 6 (0.6719), while maintaining high specificity across all classes. Its balanced accuracy, as high as 0.8949 for Class 1, reflects its ability to handle both major and minority classes effectively. In contrast, MLR and ANN show lower sensitivity, often failing to classify minority classes (0 sensitivity in multiple cases) and achieving poorer balanced accuracy overall. The findings highlight the UK as the most reliable classifier among the three, excelling in both overall and class-specific performance. Future research should focus on further optimizing the UK model and testing its applicability to other datasets and classification tasks. Additionally, expanding the dataset with more samples from underrepresented classes could further improve model accuracy.