SGRR: A two-stage method for removing raindrops from ground-based sky maps

The ground-based sky map provides real-time meteorological information for photovoltaic power stations and optimizes power generation efficiency by capturing changes in sky clouds. In order to effectively remove raindrops from the ground sky map and obtain a more accurate ground sky map, this paper proposes a new method called “Semantic Segmentation Generative Adversarial Raindrop Removal” (SGRR), which is divided into two stages: raindrop region segmentation and reconstruction. First, the ground-based sky map raindrop region segmentation dataset (GSRRS Dataset) and the ground-based sky map raindrop region reconstruction dataset (GSRRR Dataset) were constructed. During the segmentation stage, the U-shaped network was used in combination with the Depthwise Separable Receptive Field Pyramid module and Convolutional Block Attention Modules. The experimental results showed that the mean pixel accuracy, mean intersection over union, F1 score, and frames per second (FPS) on the GSRRS dataset reached 92.547%, 88.212%, 89.741%, and 120.7, respectively. In the reconstruction phase, a bidirectional interpolation filling module was proposed and combined with the optimization results of efficient channel attention mechanism. The peak signal-to-noise ratio, structural similarity, and FPS on the GSRRR dataset reached 30.23, 0.8751, and 77.6, respectively. The experiment shows that SGRR technology can accurately remove raindrops from ground-based sky images and generate images without obvious artifacts and has excellent performance.