High-quality land cover data are essential for environmental policy, spatial planning, and ecosystem monitoring. However, pixel-based classification methods, while widely used due to their practicality, often suffer from salt-and-pepper noise, which undermines map reliability. This study aimed to integrate Random Forest (RF) classification and majority filtering to enhance the quality of land cover mapping in Sukajaya District, Bogor Regency. RF was applied to Sentinel-2 image data with varying numbers of trees (ntree) to determine the optimal model performance. Subsequently, majority filtering was applied to each classification result to reduce noise and improve spatial coherence. The evaluation employed multiple accuracy metrics, including User’s Accuracy (UA), Producer’s Accuracy (PA), F1-Score, Overall Accuracy (OA), and Kappa Coefficient (KC). Comprehensive accuracy increased with the ntree until reaching an optimal point. Beyond this point, additional ntree resulted in diminishing returns. Applying majority filtering as a post-processing procedure led to further improvements in classification accuracy. While majority filtering can reduce classification noise and improve the visual quality of land cover maps, it also carries the risk of removing small, accurately classified land cover patches. This consequence is rarely discussed in similar studies. These findings highlight the importance of integrating pixel-based machine learning classification with majority filtering in land cover classification workflows, while emphasising a trade-off that tends to favour visual accuracy over the preservation of spatial detail.