Abstract: To address the challenge of accurately determining channel sandbody widths in reservoir modeling, this study presents a novel training image optimization method based on transfer learning. First, through detailed geological characterization, the parameter space for object-based modeling of sand-shale distributions was established. Six candidate training image models with varying channel widths were generated, and a representative sample set was constructed via stochastic sampling. Four pre-trained convolutional neural networks (AlexNet, ResNet50, MobileNet.V2, and Xception) were systematically evaluated, with MobileNet.V2 selected as the optimal architecture due to its lightweight design and 100% classification accuracy. Optimal sampling parameters (300 points and 200 iterations) were determined through sensitivity analysis. Well log data were then input into the transfer learning framework for hierarchical prediction, yielding 7 slices consistently classified as the 800 m channel width model. This validated candidate model 3 as the most geologically representative training image. The proposed method provides an efficient and robust workflow for training image optimization in complex reservoirs (e.g., tight oil formations), achieving significant improvements in both modeling efficiency and accuracy compared to conventional approaches.