Machine learning-based assessment of grazing-induced soil changes in Himalayan alpine grasslands

Himalayan grasslands are ecologically sensitive ecosystems facing accelerating degradation from intensive livestock grazing. This study investigated the capacity of soil physicochemical properties to classify grazing disturbance intensity in the bugyal grasslands of the Chakrata region, Uttarakhand, India. Ninety soil samples were collected during a single pre-monsoon campaign in April 2026 from three disturbance classes (non-disturbed, moderately disturbed, and highly disturbed) across two high-altitude grassland sites ( Moila Bugyal and Koti Kanasar Bugyal ), with thirty samples per class. Six parameters were measured: soil organic carbon, available nitrogen, phosphorus, potassium, bulk density, and soil pH. The data were analysed using two-way analysis of variance, Tukey’s post-hoc test, and permutational multivariate analysis of variance; three machine learning classifiers (Random Forest, Support Vector Machine, and Gradient Boosting Machine) were evaluated using five-fold cross-validation. Soil organic carbon declined most sharply from non-disturbed to highly disturbed sites (5.71% to 3.39%; -40.7%), available nitrogen fell by 34.5%, and bulk density increased by 19.5%. The disturbance gradient explained 42.5% of the total multivariate soil variance (PERMANOVA). Random Forest achieved the highest classification accuracy (86.3 ± 7.5%; kappa = 0.795), with soil organic carbon identified as the most discriminative feature (36.7% of feature importance). The results indicate that soil physicochemical properties can support machine learning-based classification of grazing disturbance in Himalayan alpine terrain.