نوع مقاله : مقاله پژوهشی
عنوان مقاله English
نویسندگان English
Predicting soil calcium carbonate equivalent (CCE) is essential for effective land management in semi-arid regions characterized by high soil and topographic variability. This study evaluated the performance of Gene Expression Programming (GEP) for spatial prediction of CCE. Environmental predictors included seven Landsat 8 spectral bands, 15 derived spectral indices, nine ASTER-derived spectral indices, and 20 topographic variables derived from the digital elevation model (DEM). Recursive Feature Elimination based on Support Vector Machines (RFE-SVM) was employed for feature selection, reducing the predictor set by 84%. The optimized GEP model, consisting of three independent genes linked by an additive function, generated an explicit and interpretable mathematical equation for CCE prediction. Model performance demonstrated moderate predictive capability, with R² and RMSE values of 0.62 and 4.13 in the training phase, and 0.52 and 4.25 in the testing phase, respectively. However, regression slopes of 0.509 and 0.487 for the training and testing datasets indicated substantial regression toward the mean and underestimation of prediction variability. The resulting prediction map revealed a distinct north–south spatial gradient consistent with regional hydrological and topographic conditions. Overall, the integrated RFE-SVM–GEP framework provides a practical basis for digital CCE mapping in semi-arid environments. Nevertheless, improving the quality and spatial representativeness of training data, incorporating more informative environmental predictors, and integrating hybrid modeling approaches are recommended to enhance prediction accuracy, particularly for extreme CCE values.
کلیدواژهها English
The precise spatial delineation of Calcium Carbonate Equivalent (CCE) in the surface horizon (0–30 cm) is a critical prerequisite for advanced pedological research and the operationalization of sustainable land management (SLM) in water-stressed, semi-arid environments. Conventional Digital Soil Mapping (DSM) techniques, particularly those leveraging opaque 'black-box' machine learning algorithms, often preclude the necessary mechanistic understanding of pedogenic factors governing CCE distribution, thereby impeding scientific knowledge transfer. This study was conceived to rigorously assess the capacity of Gene Expression Programming (GEP)—an evolutionary computation paradigm designed to yield explicit, parsimonious mathematical models—for spatially predicting CCE content. The core aim was to derive a transparent, non-linear, and fully interpretable predictive pedotransfer function within the heterogeneous Abyek Plain, Qazvin, offering a demonstrable scientific advantage over correlative, non-mechanistic models for enhanced environmental stewardship.
The study was executed within the Abyek Plain, Iran, an expansive pedoscape (approximately 60,000 ha) characterized by a semi-arid climate, thermic soil temperature regime, and diverse Quaternary geomorphic surfaces. A total of 240 georeferenced surface soil samples were acquired, ensuring representative coverage across the region's stratified pedological heterogeneity. Laboratory characterization confirmed the soil matrix as strongly alkaline and highly calcareous (mean CCE: 11.30%; pH: 7.88). The initial ancillary data catalogue comprised 51 covariates extracted at a 30-meter spatial resolution from multi-temporal median composite imagery (Landsat 8, ASTER) and an ALOS Digital Elevation Model (DEM). To address inherent multicollinearity and optimize the model’s predictive stability, the Recursive Feature Elimination (RFE) algorithm was systematically employed. RFE successfully pruned the catalogue to the optimal subset of eight predictors, including the Brightness Index, Wind Exposition, Wetness Index, DEM, Convergence Index, and PCA3. The finalized dataset was randomly partitioned into an 80% calibration set and a 20% independent validation set. The GEP framework was configured with 30 chromosomes and 6 genes, utilizing the additive fitness function to minimize the Root Mean Square Error (RMSE) and facilitate the evolution of a multi-genic prediction model.
The GEP-derived model exhibited commendable performance and robustness in spatially modeling the CCE concentration. The calibration phase yielded a Coefficient of Determination of R² = 0.62 and a minimal RMSE of 4.13%. Crucially, the model maintained its predictive efficacy on the independent validation set, achieving a statistically acceptable R² = 0.52 and an RMSE of 4.25%. This performance affirms the model's generalized applicability across the pedoscape. The explicit mathematical architecture generated by GEP facilitated the quantifiable assessment of predictor contributions:
The Brightness Index emerged as the single most influential covariate, contributing 23% of the model’s predictive power. This strong positive correlation (r=0.48) is mechanistically linked to the high spectral albedo associated with surficial salt and carbonate accumulation zones, particularly in poorly-drained, evaporative depressions. Wind Exposition followed with a substantial 18% contribution, suggesting that wind-driven sediment transport (aeolian processes) and micro-relief significantly regulate the redistribution of CCE. The Wetness Index demonstrated a significant negative correlation (r=-0.44), consistent with the hydrogeomorphic mechanism of CCE leaching in well-drained or topographically lower accumulation areas.
This investigation successfully demonstrated the utility of the GEP algorithm as a superior, mechanistically transparent alternative for the Digital Soil Mapping of CCE in complex semi-arid environments. The synthesis of RFE for predictor selection and GEP for model derivation resulted in an explicit, yet highly accurate, predictive model (R² = 0.52 validation). The resultant multi-genic pedotransfer function not only provides high-resolution spatial predictions but also offers unparalleled interpretability into the geo-environmental controls—such as spectral response linked to surficial salinization/carbonation and topographic features—that govern CCE distribution. This transparent modeling output constitutes an invaluable evidence base for precision agricultural intervention (e.g., targeted amendments, optimized irrigation scheduling) and informed environmental policy in calcareous, water-scarce regions, thereby supporting the tenets of sustainable land management. Future research should prioritize integrating time-series covariates to capture the dynamic spatio-temporal fluxes influencing CCE distribution.
The study was funded by the University of Tehran, Country Iran.
Conceptualization, Fereydoon Sarmadian; methodology, Fereydoon Sarmadian; software, Mohsen Davirand; validation, Fereydoon Sarmadian; formal analysis, Fereydoon Sarmadian; investigation, Mohsen Davirand, Sajjad Teimouri; resources, Mohsen Davirand ; data curation, Fereydoon Sarmadian; writing—original draft preparation, Mohsen Davirand; writing—review and editing, Mohsen Davirand; visualization, Mohsen Davirand; supervision, Fereydoon Sarmadian; project administration, Fereydoon Sarmadian; funding acquisition, Fereydoon Sarmadian.
All authors have read and agreed to the published version of the manuscript.
All authors contributed equally to the conceptualization of the article and writing of the original and subsequent drafts.
The authors declare that any AI and AI-assisted software was not used for writing process.
Data available on request from the authors.
The authors sincerely appreciate the support and facilities provided by University of Tehran.
The authors avoided data fabrication, falsification, and plagiarism, and any form of misconduct.
The authors declare no conflict of interest.
