Soil degradation in arid ecosystems is a major threat to sustainable development and food security, especially under accelerating climate change. Kazakhstan, where more than 70% of agricultural land suffers from salinisation, erosion, and humus loss, offers a representative case for studying climate-driven degradation. This study quantitatively assessed the influence of air temperature, precipitation, aridity index, and extreme climatic events on soil properties in the arid regions of western Kazakhstan (Atyrau and Mangystau). The analysis integrated long-term meteorological time series (1941–2023) with field and laboratory data (1967–2024) into a harmonised dataset of 1330 records. Principal component analysis (PCA) identified four degradation gradients explaining 73.6% of total variance, while Random Forest and SHAP algorithms quantified variable importance. Mean annual temperature, frequency of arid years, and aridity index were the strongest predictors of humus, salinity, pH, and CO₂ parameters, with climate factors accounting for up to 30% of soil variability. The findings demonstrate that climatic stressors are the main drivers of soil degradation in arid zones, with climate factors explaining up to 30% of the variability in key soil properties (humus, salinity, pH, and CO₂)—a substantial proportion that underscores their dominant role relative to local geochemical and anthropogenic influences. The proposed hybrid PCA—Random Forest/SHAP framework provides a robust tool for analysing climate–soil interactions and supports the design of adaptive land-use strategies to achieve Land Degradation Neutrality (LDN) in Kazakhstan and other arid regions.

Long-term soil organic carbon (SOC) monitoring provides a rare opportunity to test the performance of process-based models against real data. We evaluated RothC using 20 georeferenced soil profiles sampled in 2004 and 2024 in the semi-arid steppe of Central Kazakhstan. We used the 2004 SOC stocks for equilibrium initialization, whereas independently measured 2024 stocks were used for endpoint evaluation. A baseline-conditioned, endpoint-uncalibrated forward simulation (S1) was driven by profile-specific climate, NPP-scaled plant-C inputs, vegetation cover, and reconstructed management; a constrained inverse analysis (S2) estimated the non-negative effective C-input multiplier required to reproduce the observed endpoint, and a zero-input simulation (S3) defined the lower feasibility boundary. S1 preserved the broad between-profile variation in absolute 2024 SOC stocks (R² = 0.692), but agreement for the observed 2004–2024 SOC change was negligible (R² < 0.001; Lin’s CCC = 0.002). The direction of SOC change was reproduced at 9 of the 20 profiles, with an RMSE of 12.49 Mg C ha⁻¹ and a mean bias of −4.41 Mg C ha⁻¹. We obtained physically admissible inverse solutions for 17 profiles, with M_required ranging from 0.42 to 4.62 (mean = 2.04; median = 1.96). Profiles 295, 299, and 313 remained below the zero-input boundary under all management-timing scenarios. Together, these results show that adjusting plant-C input can reconcile modeled and observed SOC at most profiles. However, no non-negative adjustment of plant-C input could reproduce some observed endpoints. These cases indicate that other processes or sources of uncertainty should be considered.