Clastic reservoirs exhibit complex and diverse lithologies. Some lithological heterogeneities, occurring as thin but effectively low-permeability units, have pronounced impact on CO₂ flooding schemes and oil recovery. Thin low-permeability units within permeable sandbodies typically exhibit weak well-log responses, and are therefore of difficult recognition using conventional well-log analysis methods. To address this challenge, a hierarchical method is proposed for interpreting thin lithological heterogeneities by integrating wavelet transform and machine learning. The discrete wavelet transform enhances well-log responses of thin heterogeneities. An automated machine-learning framework is designed, which integrates multiple algorithms and achieves automated parameter optimization. This machine-learning method is then applied to well logs to establish a nonlinear mapping model between lithology and well-log responses. Additionally, the hierarchical nature of the workflow highlights lithological contrasts, facilitating a more accurate lithological differentiation by dividing the recognition of thin heterogeneities into three levels. Benefiting from these three advantages, the proposed method offers potential to significantly enhance the accuracy of well-log interpretations. The results demonstrate that this method yields accurate identification of lithological units as thin as 0.2 m for muddy beds and 0.3 m for diagenetic units, achieving a recognition accuracy exceeding the conventional well-log interpretations. This method also shows significant potential for broader applications, including the identification of other types of geological entities of limited thickness, and determination of reservoir parameters at fine scales.