Soil organic carbon-to-clay ratio as a proxy for land degradation: machine learning-based spatial prediction in semi-arid agricultural lands

Abstract Soil organic carbon (SOC) dynamics are central to evaluating land degradation, particularly in semi-arid regions where monitoring SOC-to-clay ratios (an indicator proposed for assessing soil resilience but still debated) remains challenging. This study employs machine learning (ML) models, including Random Forest (RF), Gradient Boosting, Classification and Regression Tree (CART) and Light Gradient-Boosting Machine (LightGBM), to spatially predict SOC-to-clay ratios across part of Şanlıurfa province, Türkiye, a semi-arid region dominated by pistachio cultivation. The study area includes Typic Calcixerepts, Calcic Haploxerepts and Typic Haplotorrerts, reflecting diverse pedological conditions. The efficacy of SOC-to-clay ratio was evaluated relative to a soil quality index (SQI) and identified texture-dependent biases. Results revealed soil texture as the dominant predictor, explaining 34-65% of variance across models, surpassing land use (7–12%). Pasturelands exhibited the highest ratios (0.21–0.47), classified as ‘very good’, due to minimal disturbance and sustained organic inputs, while croplands and pistachio systems showed ‘moderate degradation’ (≤0.26). A moderate correlation between SOC-to-clay ratio and SQI ( r = 0.51) supported its utility, though low explanatory power (R 2 = 0.26) suggests complementary indicators are needed to correct for ratio inflation in low-clay soils. Spatial predictions support EU Soil Strategy 2030 priorities, advocating for reduced tillage in croplands and perennial vegetation in pasturelands.