Using Ensemble Model Approach for Spatial Modeling of Soil Imbalanced Classes

Imbalanced data remains a widespread and significant challenge, particularly impacting machine learning algorithms. Therefore, addressing imbalanced data classification has emerged as a crucial research area within the field of data mining. This issue, often characterized by a limited number of instances in one class and a substantial number in other classes, poses substantial hurdles for machine learning algorithms. Consequently, data mining experts and machine learning professionals are actively working on refining methods and models for classifying imbalanced data with the aim of improving the accuracy of such classifications. The principal objective of this study is to precisely detect and categorize samples from the minority class, ultimately enhancing the precision of soil class classification. This research is conducted in a specific region, encompassing the southwestern territories of Zanjan province.

To achieve this objective, a total of 148 soil profiles were excavated using a regular grid pattern with an average spacing of 500 meters (and in some locations, up to 700 meters based on expert recommendations). After the samples were air-dried, they were transported to the laboratory. Physical and chemical analyses were conducted on all collected samples, including assessments of soil texture, soil pH, calcium carbonate equivalent, cation exchange capacity, electrical conductivity, organic carbon content, and gypsum content. Subsequently, the soil samples were meticulously classified and described up to the family level, following the comprehensive standards of the soil classification system. The most appropriate covariates were selected among 57 covariates including geomorphological and geological maps, digital elevation model (DEM), and data from Landsat 8 satellite images, using principal component analysis (PCA) and expert knowledge approaches for predicting soil classes selected. Saga-GIS and ENVI software were used to extract environmental covariates. Modeling of the soil-landscape relationship was performed using three algorithms, namely multinomial logistic regression (MNLR), random forest (RF), boosted regression tree (BRT) and ensemble model (after data balancing) in “R studio” software. To check the accuracy of the used model, the data was randomly divided into training and validation data. 80% of the data (118 profiles) were used for model training and 20% (30 profiles) were used as validation data for evaluation.

The results of the selection of covariates showed that 10 information covariates of geomorphological maps, geological information and features extracted from the digital elevation model (DEM), including Analytical hill shading (AHS), sunrise, valley depth (VD), LS Factor, Channel network distance (CND), Topographic wetness index (TWI) and Multi-resolution ridge top flatness (MRRTF) were selected as input variables. Based on the results of profile analysis, the soils of the region at the subgroup level were categorized into five classes, with imbalanced distribution, including Typic Calcixerepts, Typic Haploxerepts, Gypsic Haploxerepts, Typic Xerorthents, and Lithic Xerorthents. The results of evaluation metrics such as overall accuracy and Kappa index were 65% and 0.32 for the RF algorithm, %60 and 0.35 for the boosted regression tree algorithm, 65% and 0.41 for the MNLR algorithm and after balancing the data with the ensemble model approach, it was 70% and 0.62 respectively. The results of two statistics of user’s accuracy and producer’s accuracy showed that among individual models, the multinomial logistic regression model has higher accuracy in predicting soil classes. Although the ensemble model has succeeded in predicting the soil minority classes well, due to the fact that the two weaker models of the RF and BRT are involved in the modeling, It showed lower values compared to the individual multinomial logistic regression model, in predicting some classes of the majority of soil, especially the two classes of Typic Haploxerepts and Typic Xerorthents.

In summary, the results have demonstrated that when learning algorithms are individually applied, they do not exhibit high accuracy in spatially predicting soil classes. However, when these algorithms are amalgamated into an ensemble model, they exhibit remarkable accuracy in spatial soil class prediction, outperforming individual models in terms of performance and accuracy. Moreover, the ensemble model substantially enhances prediction accuracy and reduces the occurrence of misclassifications, especially at the subgroup level. While each specific model excels in predicting a particular soil classification, the cumulative ensemble models consistently outperform individual models in terms of overall performance and accuracy, underscoring the effectiveness of ensemble modeling in improving spatial soil classification.