Intelligent Bionic Polarization Orientation Method Using Improved U-Net Full-Sky Imaging Cloud Recognition Model for Cloudy Conditions

Accurate and reliable navigation is critical for the Internet of Things (IoT). Polarization-based navigation has attracted extensive attention for autonomous heading estimation in small and unmanned platforms, particularly when satellite navigation is unavailable. However, conventional polarization navigation systems suffer from degraded heading accuracy under cloudy conditions, mainly due to unstable cloud identification and insufficient feature extraction in full-sky observations. To address these issues, this paper proposes a cloud layer identification method for full-sky images based on an improved U-Net. By integrating attention and residual learning, a novel Channel Weighting-Feature Fusion (CWFF) module is embedded into the U-Net structure to strengthen deep feature representation and cloud feature memorization. Furthermore, a feature-matrix-based cloud detection scheme is developed to skip explicit imaging and improve recognition efficiency. Experimental results show that the proposed method achieves over 95% overall accuracy in cloud detection, with an average absolute error below 5%. Outdoor tests on an atmospheric polarization measurement system demonstrate that the root-mean-square error (RMSE) is reduced to 0.181°, effectively improving polarization-based heading accuracy under cloudy environments.