Geochemical-parameter-guided multimodal lithology identification for shale oil reservoirs

Reliable lithology identification in shale reservoirs is challenged by thin interlayers, heterogeneity with depth, and noise in electrical imaging logs. To address these limitations, we propose a deep-learning framework that integrates a modified You Only Look Once v5 object detector (YOLOv5) detector with K-means clustering to enhance feature extraction and recognition. The study employs a curated dataset of 2355 electrical imaging samples covering eight lithofacies, stratified to preserve class balance and validated through a combination of layered K-fold cross-validation and depth-aware sliding-window evaluation. Image quality was improved using Mosaic augmentation and Hue-Saturation-Value (HSV) perturbations, and lithological labels were cross-checked with geochemical parameters to ensure robustness. Compared with Convolutional Neural Networks (CNNs), traditional geochemical and fractal methods, and standard YOLOv5, the proposed model achieved the highest overall accuracy (97.1%) and demonstrated significant improvements in recognizing thin-layered (< 5 cm) gray mudstone. Statistical tests confirm that the improvements over baselines are highly significant (p < 0.01). Beyond accuracy, the framework provides practical interpretability through localized bounding-box overlays with confidence scores, enabling geologists to visually validate predictions on logs. This work contributes a data-driven and interpretable solution for lithology recognition, advancing the reliability of shale reservoir characterization.