ارزيابي روش‌هاي سنجش‌ازدور و نزديك در پيش‌بيني ذخيره كربن خاك با استفاده از مدل‌هاي يادگيري ماشين (منطقه مورد مطالعه: شهرستان فريدن)

eva‎luation of Remote an‎d Proximal Sensing Methods for Predicting Soil Carbon Stock Using Machine Learning Models (Case Study: Fereydun County)

جغرافياي طبيعي

Accurate quantification of soil organic carbon (SOC) an‎d soil inorganic carbon (SIC) stocks is critical for sustainable lan‎d management an‎d climate change assessments. This study aimed to eva‎luate an‎d compare the efficacy of laboratory spectroscopy (proximal sensing) an‎d remote sensing techniques within a digital soil mapping (DSM) framework for predicting an‎d spatially delineating SOC an‎d SIC in Fereydun County. A total of 100 soil samples were collected using a Latin hypercube sampling design, an‎d machine learning algorithms, including Ran‎dom Forest (RF) an‎d Support Vector Machine (SVM), were employed for predictive modeling.Spectral modeling results (point-scale prediction) revealed that optimal spectral preprocessing strategies vary between SOC an‎d SIC. The highest predictive accuracy for SOC was achieved with the SVM model combined with first-derivative preprocessing (R² = 0.65, RPD = 1.7), whereas SIC prediction was best modeled using RF with detrending preprocessing (R² = 0.90, RPD = 3.1). At the spatial mapping scale, hybrid models integrating remotely sensed data (satellite imagery) an‎d environmental covariates with spectrally predicted values demonstrated superior performance. Specifically, SVM emerged as the optimal model for SOC mapping (R² = 0.81), while RF was most effective for SIC mapping (R² = 0.84).The findings indicate that although laboratory spectroscopy alone provides high point-scale predictive accuracy, the integration of proximal an‎d remote sensing data within a DSM framework represents the most robust approach. Rather than being competing methods, proximal an‎d remote sensing are complementary technologies, an‎d their synergistic application produces highly accurate an‎d reliable soil carbon maps, highlighting the limitations of direct prediction from satellite data alone.

5

147689