This study aims to quantify aboveground carbon at risk in semi-arid mountain forests of the Moroccan High Atlas and to test whether composite vulnerability is higher in structurally dense stands than in open stands. A 30-year Landsat time series (1995–2025) was analysed using LandTrendr applied to the NBR index, combined with probabilistic field inventory data and penalized logistic regression. A composite vulnerability index was developed by integrating disturbance probability, spectral severity, and recovery capacity, with climatic predictors including precipitation seasonality, Heat Load Index, and 12-month SPEI. Total aboveground carbon stock was estimated at 122,474 tC, of which 15,985 tC (13.05%; 95% bootstrap CI: 11,654–20,293) is classified as at risk. A pronounced hotspot in dense forest strata accounts for 40.4% of carbon at risk while covering 26.2% of the area. After accounting for collinearity with elevation, precipitation seasonality emerged as the only significant predictor of disturbance probability (β = 1.184; p = 0.002; AUC = 0.827), providing partial support for the hypothesised climatic control on disturbance. Field validation against mortality observations showed perfect sensitivity (1.000) and moderate specificity (0.478), indicating a tendency toward early detection of partial dieback rather than omission of true events. The results suggest that structurally dense stands concentrate a disproportionate share of carbon vulnerability, consistent with a climatic maladaptation mechanism linked to past establishment conditions under more favourable moisture regimes. However, estimates should be interpreted as conservative upper bounds due to moderate model specificity (0.478), uncertainty in spectral proxies for disturbance and recovery, and limitations in stratum-level biomass generalisation. The framework provides a reproducible approach for mapping carbon at risk by integrating remote sensing, field inventories, and climate data, and can support prioritisation of monitoring and management interventions such as targeted thinning and regeneration assistance in high-risk dense stands. The main contribution lies in delivering a transferable methodology for quantifying spatial carbon vulnerability in semi-arid forest ecosystems where empirical data remain scarce.