As a strategically critical resource, bauxite exploration requires urgent advancement. To address the limitations of traditional prospecting methods in efficiency and accuracy, this paper takes the karstic bauxite in the Pingguo area of western Guangxi as the research object and proposes a deep learning- based prospecting prediction method driven by multi- source data fusion. Based on the U- Net benchmark model, the study focuses on the impact of constructing multi- source geoscience data training sets using the sliding window technique and optimizing network architecture parameters on model performance. It also quantitatively evaluates model performance by introducing a composite score S (weighted by the intersection over union IoU, F1 score, and normalized training time Tn) and subsequently constructs a deep learning- based prospecting prediction model. The results show that: ① When constructing the dataset using sliding window technology, the S- value with a 75% overlap rate increased significantly by 71. 04% compared to that with a 0% overlap rate, indicating that this setting can effectively enhance the data and significantly improve model performance; ② In model architecture optimization, the optimal combination was determined through controlled experiments (64 base channels, replicate padding, SE+ module, ELU activation function, and Cross- Entropy+Dice composite loss function), which further increases the S- value by 8. 72%, significantly improving the recognition capability for complex geological features. Finally, prospecting prediction was conducted in the target area combined with multi- source data, successfully delineating five prospecting targets. The integration of multi- source data and deep learning technology enriches the prospecting prediction theory for karstic bauxite and provides scientific basis for exploration of this deposit type.