Abstract: The neighborhood median pixel method is a pixel-based classification technique used in remote sensing to categorize Landsat 8 imagery into three distinct classes: vegetation, water, and built- up areas. This method utilizes the median value in a pixel’s neighborhood to generate a lookup table and uses a score profile that assigns score-level values for eventual classification. Initially proposed using a 9 × 9 neighborhood size, later research exploring different neighborhood sizes found a 13 × 13 configuration achieved better overall accuracy. A follow-up study then employed particle swarm optimization to improve the arbitrary point values used in the score profile. Building on these developments, this study thoroughly analyzes the key principles of the algorithm and its operational behavior in the classification procedure, focusing on the structure and computation of the score values of its score profile. The analysis demonstrates that the score profile exhibits invariance under translation (adding a constant to all score values) and scaling (multiplying all score values by a positive constant), indicating that such transformations do not affect the resulting class predictions. Recognizing this, the study reformulated the score profile to leverage the invariance to reduce the dimensions of the search space, allowing the optimization algorithm to search more efficiently. By reducing the tunable parameters – anchoring one score at 0 and normalizing another to 1 – the dimensionality of the search space is reduced, improving computational efficiency. Despite this simplification, the classification accuracy marginally increased, from 94.75 to 95.25% (9 × 9 neighborhood) and from 96 to 97.50% (13 × 13 neighborhood). These gains suggest that the initial configuration is already stable and performs reliably. In summary, the original score profile is sufficient to achieve effective classification, though targeted optimization of the scores can deliver modest improvements in accuracy.