ResAG-UNet: A Novel Residual Attention Gated UNet for Cloud Segmentation in Sky Image

Cloud cover significantly impacts the solar radiation reaching the Earth's surface, thereby influencing the efficiency and output of solar energy systems. Consequently, an accurate cloud segmentation approach is crucial for understanding fluctuations in solar irradiance in real time and future ahead. Such understanding aids in optimizing energy production and grid management. In this article, we designed a novel deep learning architecture called Residual Attention Gated-UNet (ResAG-UNet) for accurate cloud segmentation. The proposed ResAG-UNet integrates residual blocks in both the encoder and decoder paths, along with an attention mechanism in the decoder path. The inclusion of residual blocks facilitates faster gradient movement due to skip pathways across them, thereby enhancing training efficiency. Furthermore, the incorporation of an attention module in ResAG-UNet allows for the learning of attention coefficients for various pixels. This mechanism actively highlights crucial characteristics while suppressing less significant ones in the cloud image. The proposed ResAG-UNet model is assessed and compared with benchmark segmentation models using NITK and SWIMSEG sky datasets. The proposed approach yields mean IOU, precision, recall, F1 score, accuracy of (0.8616, 0.8826), (0.9761,0.9965), (0.9863,0.9764), (0.9237,0.9613), and (0.9424, 0.9651) on the NITK and SWIMSEG sky datasets, respectively.