Alpine agroecosystems face fertilizer-induced nutrient imbalances, including early nitrogen accumulation and late-season nitrate residue. While organic substitution is promoted as a sustainable practice, different substitution ratios may generate contrasting responses among nutrient availability, rhizosphere function, and SOC temporal stability. We established a field experiment in Tibetan highland barley with five treatments at equivalent N, P, and K rates: 100% NPK, 75% NPK + 25% cattle manure (CM), 50% NPK + 50% CM, 25% NPK + 75% CM, and 100% CM, plus an unfertilized control. Soils were sampled from 0–10 cm bulk soil, 10–20 cm bulk soil, and rhizosphere soil (0–10 cm) at booting, filling, and maturity. Stoichiometric ratios in the 0–10 cm layer changed mainly at maturity, whereas responses in the 10–20 cm layer and rhizosphere were weaker, indicating stage- and compartment-dependent treatment responses. The treatments showed three response modes: mineral fertilizer-dominated nutrient pulses, manure-dominated late-season nitrate residue and surface SOC stability, and partial substitution patterns with comparatively balanced nutrient responses. No single substitution ratio optimized all measured indicators. We therefore propose these modes as a framework for organizing treatment-specific nutrient responses and as candidates for further multi-year and multi-site testing, rather than as ready-to-apply fertilization prescriptions.

Soils underpin terrestrial ecosystem functioning and exert strong control over farmland productivity, nutrient turnover, environmental quality, and agricultural sustainability. Reliable prediction of soil nutrients and pH is therefore important for precision fertilization, soil-quality assessment, and the sustainability of farmland management, particularly in fragile alpine agroecosystems. However, the comparative performance of candidate algorithms for estimating soil properties in alpine croplands of the Qinghai-Tibetan Plateau remains poorly characterized. Here, we assessed nine approaches, namely random forest (RF), generalized boosting regression (GBR), multiple linear regression (MLR), artificial neural network (ANN), generalized linear regression (GLR), conditional inference tree (CIT), extreme gradient boosting (eXGB), support vector machine (SVM), and recursive partitioning regression tree (RRT), for predicting soil carbon, nitrogen, phosphorus, potassium, and pH from annual mean temperature, annual precipitation, annual solar radiation, and maximum normalized difference vegetation index (NDVImax). Field observations were obtained from 103 sampling plots, with four individual soil samples collected per plot and depth, resulting in 412 observations for each depth layer. Clear contrasts in model performance emerged under alpine agricultural conditions. Across all soil variables and depths, eXGB and RF showed the best overall performance, with the lowest mean relative bias values of 1.35% and 1.55%, respectively, and the lowest mean RMSE values of 0.43 and 0.47. These results indicate that nonlinear ensemble algorithms better captured the heterogeneous relationships between environmental covariates and soil properties than linear models or single-tree approaches. Overall, RF and eXGB provide a practical methodological basis for precision fertilization, soil-quality assessment, and long-term sustainability in fragile alpine agroecosystems.