Accurate and real-time prediction of ship infrared (IR) radiation characteristics is crucial for maritime search-and-rescue (SAR) operations and naval stealth management, but traditional methods are either computationally inefficient (e.g., Computational Fluid Dynamics requiring hours per scenario) or lack cross-scenario generalization (e.g., pure data-driven models). This study proposes a Kernel-Attention Mechanisms U-Net (KEU-Net) framework to address these issues. The framework integrates U-Net’s robust spatial feature extraction—via skip connections that preserve edge details of small IR targets (e.g., exhaust plumes, hull thermal gradients)—with a kernel-attention module. This module dynamically adjusts feature weighting based on maritime environmental factors (solar angles, sea states) and ship operational parameters (heading, speed), avoiding parameter redundancy of multi-layer perceptions (MLPs). Trained on three datasets covering seasons, weather, and thermal suppression scenarios, KEU-Net achieves 97.3% accuracy against finite element benchmarks, maintains ≤1.7°C temperature error across Beaufort 3–8 sea states, and completes full-azimuth IR pattern prediction in 12 seconds per scenario (32× faster than response surface models). It supports real-time SAR and stealth planning, with strong compatibility with naval systems.