An effective implementation and assessment of a random forest classifier as a soil spatial predictive model

Mapping the spatial distribution of soil classes is important for informing soil use and management decisions. This study aimed to effectively implement Random Forest (RF) model and to evaluate the behaviour and performance of the model for soil classification of Indian districts. Soil-forming factors, known as ‘scorpan,’ are selected as environmental covariates to tune RF model to classify 11 different soil categories. Thirty-five digital layers are prepared using different satellite data ALOS (Advanced Land Observing Satellite) digital elevation model, Landsat-8, Moderate Resolution Imaging Spectroradiometer normalized difference vegetation index product, RISAT-1 (Radar Imaging Satellite-1), Sentinel-1A] and climatic data (precipitation and temperature) to represent scorpan environmental covariates in the study area. The RF parameters corresponding to highest Cohen’s kappa coefficient (κ) value and lowest number of random split variables are considered optimum values for RF model. Model behaviour evaluation is based on mapping accuracy, sensitivity to data set size, and noise. Two other machine-learning methods, CART (Classification and Regression Tree) decision tree (CDT) and CART ensemble bagger (CEB), are used to provide the comparative study. To access behaviour of models to the false data set, noise in training set is produced by assigning a false class to the training set in 5% increment. Comparative performance of RF model is based on quality assessment measures. To evaluate the performance of models, marginal rates, F-measure, and Jaccard’s coefficient of the community, classification success index and agreement coefficients are selected under quality assessment measures. The score is calculated to rank the algorithm. RF model shows high stability against data set reduction in comparison to other methods. The results show that the abrupt change in accuracy is only observed after 60% training data reduction in RF model; however, significant decrease in accuracy can be noted after 45% and 25% data reduction in CEB and CDT, respectively. The RF model shows comparatively the greater resistance to noise. Overall, RF model has performed better than CDT and CEB to classify soil categories in the study area. The results of this research provide new insights into the performance of RF in the context of soil class mapping.