The sandstone reservoirs in the Doseo Basin, Chad, developed under a strike-slip tectonic setting, exhibit complex oil–water relationships with unclear resistivity contrasts between oil and water zones, and overlapping oil–water signals in NMR T2 distribution. This study proposes a two-stage machine learning method based on clustering analysis and classification prediction. By utilizing well testing results and 2D NMR data to expand labeled data, and applying PCA for dimensionality reduction on T2 distribution as feature logs, the multi-resolution graph-based clustering (MRGC) method is used for clustering analysis. The clustering results are labeled to form a classification learning sample dataset, and the K-nearest neighbors (KNN) method is employed to predict fluid types in new wells. This method accurately reflects changes in lithology and flushed zone fluid properties, identifying four types of oil zones and six types of water zones with an accuracy of 89.3%. Validation through a tested well demonstrates the method's effectiveness. This approach considers the relatively abundant and reliable 1D NMR data, collecting additional 2D NMR data only in some key wells, thus saving exploration costs and improving the timeliness and accuracy of well log interpretations.