In maritime transportation and port handling operations, rainfall poses a threat to moisture-sensitive cargoe, leading to potential mildew and economic losses. While many vessels are equipped with Closed-Circuit Television (CCTV) to monitor operations, they generally lack professional meteorological sensors, which renders rainfall assessment an inefficient manual process. To address this, we developed a deep learning framework based on EfficientNetV2-S for automated rainfall identification and rain intensity classification. We propose a Rain-Sensitive Attention (RSA) module that incorporates physics-inspired directional priors to perceive the anisotropic geometry of rain streaks. Unlike conventional isotropic attention mechanisms, RSA utilizes multi-branch elongated kernels to explicitly decouple high-frequency rainfall textures from complex maritime backgrounds, such as water accumulation and surface reflections. The RSA module is strategically deployed in both shallow and deep layers, enabling a hierarchical feature evolution from local rain-streak cues to global precipitation patterns. For the rain intensity classification task, a transfer learning strategy is adopted to share the underlying visual features of rain patterns. This improves recognition accuracy, particularly for light rain, where rain intensity data in maritime environments is scarce. Furthermore, the integration of Consistent Rank Logits (CORAL) ordinal learning and a threshold-gated decoding mechanism ensures the monotonic consistency of prediction results. Experimental results show that in the rainfall identification task, the model achieves an accuracy of 95.23%. In the rain intensity classification task, it attains an accuracy of 93.26% with a Quadratic Weighted Kappa (QWK) coefficient of 95.87%.
Wang et al. (Mon,) studied this question.