Soil salinity is a major environmental hazard in arid regions, directly affecting soil and water resource productivity and ecosystem sustainability. This study aimed to develop and optimize an ensemble machine learning model for improving the spatial prediction of soil salinity and identifying the key environmental factors controlling its distribution in the Mahvelat Plain, Khorasan Razavi Province, Iran. A total of 137 soil samples were collected, and electrical conductivity of the saturated soil extract (ECe) was measured. In addition, 32 auxiliary variables, including remote sensing spectral indices, topographic features, geological information, land use data, and groundwater quality parameters, were used in the modeling process. Four machine learning models, including Random Forest (RF), Support Vector Machine (SVM), Boosted Regression Tree (BRT), and Generalized Linear Model (GLM), along with a combined model, were developed and evaluated. The results showed that the combined model outperformed the individual models, achieving the highest prediction accuracy (R² = 0.81, RMSE = 5.14, and MAE = 4.11). Among the individual models, SVM showed the best performance (R² = 0.76). Highly saline areas were mainly concentrated in the central, western, and northwestern parts of the study area, particularly around Playa Bajestan, with an overall increase in salinity from east to west. Groundwater quality parameters, especially HCO₃⁻, SAR, and TDS, were the most influential predictors. The combined model also showed a greater ability to estimate extreme salinity values, reaching 21 dS/m. Overall, the proposed framework provides an accurate and reliable approach for soil salinity mapping and sustainable soil resource management.
Introduction
Nowadays, soil salinity has become an escalating environmental crisis that directly threatens global food security. In arid regions, the gradual accumulation of salts degrades both the physical and chemical properties of soil, posing a serious obstacle to agricultural production. The Mahvelat Plain in eastern Iran is one such area. It is an important agricultural zone well known for its intensive pistachio cultivation. However, the region’s hot and dry climate, together with its proximity to the Bajestan playa, creates conditions that strongly favor salinization. The present study attempts to bridge the gap between traditional soil sampling and predictive digital soil mapping using an advanced ensemble machine learning approach. According to official reports, 424 million hectares of topsoil and 833 million hectares of cultivated soils in the world are affected by salt. Despite the increasing availability of remote sensing and environmental covariates, accurately characterizing the spatial heterogeneity of soil salinity in arid agricultural landscapes remains challenging because salinity is controlled by complex interactions among soil, groundwater, topography, climate, and land-use conditions. Therefore, integrating multiple machine learning algorithms within an ensemble framework may provide a more robust approach for improving prediction accuracy and reducing the limitations associated with relying on a single predictive model.
Material and Methods
Soil salinity values have previously been assessed through laboratory analyses, which have often been time-consuming and costly. Due to the limitations of past scientific methods, alternative techniques such as geostationary and remote sensing are currently being used to predict soil salinity in unsampled areas. Also, the integration of digital soil mapping techniques and machine learning to estimate soil properties has also expanded the application of these methods in soil salinity prediction. The study was conducted in the Mahvelat Plain. A total of 137 soil samples were collected to determine the electrical conductivity of the saturated paste extract (ECe). To account for spatial complexity, 32 environmental covariates were selected. These included spectral remote sensing data, terrain attributes derived from a digital elevation model (DEM), geological maps, and groundwater chemical analyses. Four different models were trained initially: random forest (RF), support vector machine (SVM), boosted regression trees (BRT), and generalized linear model (GLM). Afterwards, a stacking ensemble model was developed using a five‑fold cross‑validation framework to generate the final salinity prediction map.
Results and Discussion
Model evaluation of model performance clearly showed that the stacking ensemble approach was superior. This model achieved high statistical accuracy, with an R² of 0.81 and an RMSe of 5.14, thereby outperforming each of the individual RF, SVM, BRT, and GLM models. The spatial distribution map revealed a clear salinity gradient, with values increasing progressively from the eastern part toward the western and northwestern parts of the plain. According to the variable importance analysis, soil salinity in the study area is mainly governed by subsurface hydro geochemical conditions. In particular, the strong correlations observed for HCO₃ ‑SAR and TDS indicate that groundwater quality is the primary factor driving soil degradation in this plain. Unlike single models, the ensemble framework proved notably more capable of capturing extreme salinity, which is of particular concern for agricultural management. The ensemble approach outperformed all individual models, achieving an R² of 0.81, RMSe of 5.14 dS m⁻¹, and MAE of 4.11 dS m⁻¹. The final salinity map revealed high values in the western and northwestern parts of the plain near the Bajestan playa. Variable importance analysis identified groundwater hydro geochemical parameters, particularly bicarbonate (HCO₃⁻), sodium adsorption ratio (SAR), and total dissolved solids (TDS), as the primary predictors. Notably, the ensemble model substantially improved the estimation of extreme salinity values exceeding 21 dS m⁻¹ compared to any single model.
Conclusion
The results confirmed that ensemble machine learning techniques can markedly improve the accuracy of soil salinity mapping in arid regions. By capturing the nonlinear interactions between groundwater quality and soil surface properties, the proposed framework offers a practical and reliable tool for site-specific soil management. For pistachio farmers and regional policymakers in Iran, this tool can help identify priority areas for drainage and desalination interventions, thereby supporting more sustainable agricultural development in salt-affected environments. Overall, the proposed approach can serve as an efficient tool for the intelligent management of water and soil resources in saline lands, particularly in pistachio‑growing regions of the country. On the other hand, the results of this research showed that despite the extreme spatial and temporal variability of salinity, the use of remote sensing data and auxiliary variables in the framework of machine learning can predict soil salinity patterns with optimal accuracy. The resulting spatial information can also provide a scientific basis for prioritizing field investigations, optimizing irrigation and drainage practices, and allocating soil reclamation resources to areas with the greatest salinity risk. Furthermore, the proposed framework is transferable to other salt-affected arid and semi-arid agricultural regions where sufficient soil observations and environmental covariates are available, providing a scalable approach for supporting sustainable land and water management.