نوع مقاله : پژوهشی
نویسندگان
1 استادیار، بخش مهندسی منابع طبیعی و محیط زیست، دانشکده کشاورزی، دانشگاه شیراز، شیراز، ایران
2 محقق پسادکتری،گروه مدیریت مناطق خشک و بیابانی، دانشکده منابع طبیعی و محیط زیست، دانشگاه فردوسی، مشهد، ایران
3 گروه مدیریت مناطق خشک وبیابانی، دانشکده منابع طبیعی و محیط زیست دانشگاه فردوسی مشهد، مشهد، ایران
4 دانشجو دکتری، گروه مهندسی آب، دانشکده کشاورزی، دانشگاه شهرکرد، شهرکرد، ایران،
5 گروه مهندسی طبیعت، دانشکده منایع طبیعی، مجتمع آموزش عالی سراوان، سراوان، ایران،
چکیده
کلیدواژهها
موضوعات
عنوان مقاله [English]
نویسندگان [English]
Introduction and Problem Statement
Soil erosion represents one of the most critical environmental challenges facing sustainable land management in Iran's Golestan Province. This comprehensive study addresses the pressing need for accurate erosion prediction models in the vulnerable ecosystems of Maraveh Tappeh and Kalaleh counties, where water erosion has significantly impacted agricultural productivity, natural resource conservation, and ecosystem stability. The research responds to the growing concern over land degradation in these regions, where conventional erosion assessment methods have proven inadequate for capturing the complex interplay of environmental factors driving soil loss.
Methodology Framework
The research employed an integrated approach combining remote sensing technology, geographic information systems (GIS), and sophisticated machine learning algorithms to develop a robust predictive model for water erosion susceptibility. The methodological framework encompassed four distinct phases: data acquisition and preprocessing, feature selection and optimization, model development and training, and validation and accuracy assessment.
Data Collection and Processing
The study incorporated eight critical environmental and climatic indices as predictive variables, carefully selected based on their documented influence on erosion processes. These included: Precipitation patterns derived from meteorological station data and satellite-based precipitation products Vegetation cover density using the Normalized Difference Vegetation Index (NDVI) Soil characteristics through the Topsoil Grain Size Index (TGSI) Land use and land cover classifications from high-resolution satellite imagery Surface moisture content via the Normalized Difference Moisture Index (NDMI) Topographic features from Digital Elevation Model (DEM) data Bare soil exposure using the Bare Index (BI) Soil salinity levels through the Normalized Difference Salinity Index (NDSI) Data were sourced from multiple platforms including Landsat 8 OLI/TIRS sensors, SRTM DEM, and local meteorological stations, ensuring comprehensive spatial and temporal coverage. All datasets underwent rigorous preprocessing, including atmospheric correction, geometric rectification, and resolution standardization to 30-meter pixels.
Machine Learning Implementation
Three state-of-the-art machine learning algorithms were implemented and rigorously compared: Boosted Regression Trees (BRT): Selected for their ability to handle complex nonlinear relationships and automatic feature selection capabilities Random Forest (RF): Chosen for its robustness against overfitting and effectiveness with high-dimensional data Support Vector Machine (SVM): Implemented for its strength in managing small sample sizes and high-dimensional feature spaces The modeling process incorporated advanced techniques including k-fold cross-validation (k=10) and hyperparameter optimization to ensure model robustness and generalizability.
Ensemble Model Development
A novel ensemble approach was developed by integrating the strengths of individual models through weighted averaging. This sophisticated combination mechanism accounted for the relative performance of each base model, giving higher weights to more reliable predictors while mitigating individual model weaknesses.
Results and Performance Metrics
The ensemble model demonstrated exceptional predictive capability, achieving remarkable validation scores: Receiver Operating Characteristic (ROC): 0.93 Kappa Coefficient: 0.90 True Skill Statistic (TSS): 0.89 These results significantly outperformed all individual models, with the ensemble approach reducing prediction error by approximately 23% compared to the best-performing single model.
Spatial Analysis and Risk Zonation
The final erosion susceptibility map revealed distinct spatial patterns, identifying northern, western, and southwestern sectors as high-risk zones. These areas exhibited strong correlations with specific environmental characteristics: Slope gradients exceeding 15% NDVI values below 0.3 indicating sparse vegetation cover Annual precipitation intensity above regional averages Predominance of erosive soil types and unsustainable land use practices
Factor Importance Analysis
Comprehensive sensitivity analysis identified precipitation intensity and vegetation cover density as the dominant factors influencing erosion susceptibility, collectively accounting for 64% of the model's predictive power. Secondary factors included slope characteristics (18%) and soil properties (12%), while land use patterns contributed approximately 6% to the model's output.
Practical Applications and Policy Implications
The research outcomes provide valuable tools for: Land Use Planning: Enabling targeted conservation efforts in high-priority areas Resource Allocation: Facilitating efficient deployment of soil conservation resources Policy Development: Supporting evidence-based environmental policy formulation Risk Management: Assisting in the design of early warning systems for erosion prevention
Comparative Advantages and Methodological Contributions
This study demonstrates significant advancements over traditional erosion assessment methods through: Enhanced prediction accuracy through ensemble learning Improved spatial resolution enabling parcel-level assessment Robust validation through multiple statistical metrics Comprehensive factor integration capturing complex environmental interactions
Conclusion and Future Research Directions
The research successfully establishes ensemble machine learning as a powerful paradigm for erosion susceptibility mapping in semi-arid environments. The methodology offers scalable solutions adaptable to similar ecological contexts globally. Future research directions should focus on: Incorporating climate change projections for long-term erosion forecasting Integrating real-time monitoring data for dynamic model updating Expanding the model to account for anthropogenic factors and land management practices Developing multi-temporal analyses to capture seasonal and inter-annual variations This comprehensive approach represents a significant step forward in sustainable land management, providing scientists, policymakers, and land managers with reliable tools for combating soil erosion and promoting environmental conservation in vulnerable ecosystems. The research methodology and findings contribute substantially to the global effort to achieve land degradation neutrality and sustainable development goals.
کلیدواژهها [English]
)