Comparison of Machine Learning-Based Prediction of Qualitative and Quantitative Digital Soil-Mapping Approaches for Eastern Districts of Tamil Nadu, India

Abstract

The soil–environmental relationship identified and standardised over the years has expedited the growth of digital soil-mapping techniques; hence, various machine learning algorithms are involved in predicting soil attributes. Therefore, comparing the different machine learning algorithms is essential to provide insights into the performance of the different algorithms in predicting soil information for Indian landscapes. In this study, we compared a suite of six machine learning algorithms to predict quantitative (Cubist, decision tree, k-NN, multiple linear regression, random forest, support vector regression) and qualitative (C5.0, k-NN, multinomial logistic regression, naïve Bayes, random forest, support vector machine) soil information separately at a regional level. The soil information, including the quantitative (pH, OC, and CEC) and qualitative (order, suborder, and great group) attributes, were extracted from the legacy soil maps using stratified random sampling procedures. A total of 4479 soil observations sampled were non-spatially partitioned and intersected with 39 environmental covariate parameters. The predicted maps depicted the complex soil–environmental relationships for the study area at a 30 m spatial resolution. The comparison was facilitated based on the evaluation metrics derived from the test datasets and visual interpretations of the predicted maps. Permutation feature importance analysis was utilised as the model-agnostic interpretation tool to determine the contribution of the covariate parameters to the model’s calibration. The R2 values for the pH, OC, and CEC ranged from 0.19 to 0.38; 0.04 to 0.13; and 0.14 to 0.40, whereas the RMSE values ranged from 0.75 to 0.86; 0.25 to 0.26; and 8.84 to 10.49, respectively. Irrespective of the algorithms, the overall accuracy percentages for the soil order, suborder, and great group class ranged from 31 to 67; 26 to 65; and 27 to 65, respectively. The tree-based ensemble random forest and rule-based tree models’ (Cubist and C5.0) algorithms efficiently predicted the soil properties spatially. However, the efficiency of the other models can be substantially increased by advocating additional parameterisation measures. The range and scale of the quantitative soil attributes, in addition to the sampling frequency and design, greatly influenced the model’s output. The comprehensive comparison of the algorithms can be utilised to support model selection and mapping at a varied scale. The derived digital soil maps will help farmers and policy makers to adopt precision information for making decisions at the farm level leading to productivity enhancements through the optimal use of nutrients and the sustainability of the agricultural ecosystem, ensuring food security.