Urban expansion in rapidly developing regions like Haldwani, Uttarakhand, necessitates accurate land cover monitoring to support sustainable planning. This study evaluates machine learning (ML) algorithms Random Forest (RF), Support Vector Machine (SVM), and XGBoost for land cover classification using Sentinel-2 multispectral data (B2, B3, B4, B8) and NDVI in R Studio. Five land cover classes (built-up, barren land, crops, forest, water) were classified using 200 manually annotated samples per class. XGBoost demonstrated superior performance with 95.9% overall accuracy and a Kappa coefficient of 0.9588, indicating almost perfect agreement. RF achieved 93.5% accuracy (Kappa = 0.92), while SVM showed 92.5% accuracy (Kappa = 0.9063). The high Kappa values confirm robust class separability across all models. XGBoost excelled particularly in built-up area identification (96.5% F1-score), significantly reducing confusion with spectrally similar barren lands. The results highlight: (i) Sentinel-2’s effectiveness for detailed land cover mapping in heterogeneous landscapes, (ii) XGBoost’s advantage in handling complex urban spectral patterns, and (ii) the methodology’s potential for operational land-use monitoring. This approach provides urban planners and environmental agencies with a cost-effective tool for tracking land-use changes and supporting evidence-based decision-making in rapidly urbanizing Himalayan regions. The integration of open-source satellite data with optimized ML algorithms offers scalable solutions for sustainable development challenges in mountainous terrain.