<![CDATA[Land cover conversion dynamics in Penajam Paser Utara (PPU) Regency have accelerated drastically following its designation as the core and buffer zone of the Nusantara Capital City (IKN). Optical remote sensing monitoring using Sentinel-2 imagery faces severe physical constraints from tropical atmospheric interference, haze, and pronounced spectral confusion between natural rainforest canopies and oil palm plantations exhibiting overlapping chlorophyll biomass reflectance. This study aims to improve the accuracy and robustness of 6-class land cover classification (Forest, Cultivated Vegetation, Settlement, Water Body, Bare Land, and Oil Palm) by integrating Histogram Specification (HistSpec) radiometric normalization, feature engineering of 9 spectral attributes, and Random Forest (RF) ensemble classification. The performance of RF was systematically benchmarked against six alternative Machine Learning algorithms (Extra Trees, Decision Tree, Gradient Boosting, Logistic Regression, Naive Bayes, and SGD Classifier) using 675 points (472 training and 203 independent testing points). Experimental findings demonstrate that Histogram Specification coupled with Random Forest attained stable performance, Overall Accuracy (OA) of 79.80% (an improvement from 78.33% on the RAW scenario) and a Cohen's Kappa coefficient of 0,7549. In contrast, Gradient Boosting exhibited severe overfitting on raw imagery (accuracy plummeted from 83.25% on RAW to 78.33% on HistSpec). Given its high computational efficiency (~123 ms) and scale-invariance, Random Forest proves to be the most dependable model for generating distortion-free thematic land cover maps to support sustainable spatial planning in the IKN region.]]>
