Satellite image classification plays a critical role in remote sensing by enabling accurate land use analysis. This study investigates the effectiveness of traditional machine learning (ML) techniques by extracting and integrating handcrafted features from local descriptors, texture patterns, and statistical color features across seven color spaces. These features are evaluated using eight different ML methods for land use classification on the UCMerced dataset. Experimental results demonstrate that combining local descriptors, texture-based features, and statistical color information yields a peak accuracy of 91.67% using the CatBoost classifier with 263-dimensional feature vectors–closely matching the performance of ResNet50 (91.72%) while requiring significantly less computational overhead. The study further reveals that the optimal feature dimensionality lies within the range of 105 to 300, balancing discriminative power and redundancy. These findings emphasize the enduring value of traditional machine learning approaches in satellite image classification, particularly in scenarios where interpretability and computational efficiency are essential.