<p>In this study, we propose a two-step binarization method for hand-drawn architectural floor plans to transform them into usable formats for indoor spatial modeling. First, a Gaussian Mixture Modeling (GMM) was adopted to remove texture-like noise from the background. Second, a Random Forest (RF) model was trained using 24 features to classify the diverse pixel values in the image. The model’s key task was to identify whether each pixel represented an interior object (foreground) or the stained, degraded background. Feature importance analysis identified the mean of pixel values, central pixel value, and slice transform matrix as critical and complementary features, enhancing robustness across various degradation levels and architectural complexities. Our machine learning-based binarization method successfully separated meaningful architectural elements from noise, leaving only the clean architectural components against a plain background. Moreover, to assess the model’s ability to generalize, the model was applied to an entirely separate architectural drawing dataset previously unseen by the model during training. The experimental results demonstrated that our proposed method significantly outperformed existing binarization techniques on both the test dataset and a hand-drawn architectural drawing dataset from an entirely different domain. Specifically, on the test dataset, our method achieved a precision of 0.985, a recall of 0.990, an accuracy of 0.976, and an F1-score of 0.987. When evaluated on the unseen distinct dataset (VERSPERA dataset), the proposed method attained a precision of 0.998, a recall of 0.968, an accuracy of 0.970, and an F1-score of 0.983. These superior performance metrics underscore the robust generalization capabilities of our method across diverse datasets, with performance that even exceeds that of the state-of-the-art deep learning techniques.</p>