For Visible-Infrared Person Re-Identification (VI-ReID) tasks, current mainstream data augmentation methods primarily focus on color channel enhancement, failing to address common challenges like occlusion and viewpoint variations. To enhance network generalization capability and extract more robust cross-modality invariant features, we propose Diversity-Enhanced Robust Information Mining Network (DRIM-Net) incorporating two innovative data augmentation methods: Proportional Pedestrian Crop (PPC) and Cross-Modal Random Occlusion Augmentation (CMROA). PPC addresses the human body proportion distortion caused by conventional cropping strategies through an optimized approach. CMROA enhances model robustness in complex scenarios by simulating realistic occlusion patterns. Furthermore, we design Diversity-Enhanced Multi-level Feature Approach (DEMFA) to guide the network in learning modality-invariant information while incorporating multi-level diversity information to strengthen feature mining capabilities. Experimental results demonstrate that DRIM-Net achieves superior performance across multiple benchmark datasets including SYSU-MM01, RegDB, and LLCM, outperforming existing state-of-the-art methods.