Dual-Spectrum All-Sky camera Cloud Classifier by means of Computer Vision Models

This paper deals with the problem of automatically classifying clouds, comparing the performance of different neural networks using infrared-only images, visible-only images, and visible plus infrared images through constructing a labeled dataset from ground-based all-sky camera images. This work is particularly valuable for researchers in renewable energy forecasting, atmospheric science, and computer vision, as well as for applications in grid management and climate monitoring. In the context of photovoltaics, accurate cloud type classification supports nowcasting in microgrids, enabling more reliable shortterm predictions of solar power generation from photovoltaic plants. The images are pre-processed to remove the geometrical distortion introduced by the fisheye lenses of the cameras and then fed to an EfficientNet-based model; two possible ways of combining the two types of images are explored, combining the feature vectors at two different steps. The performances of the three approaches are then compared. The results show that the models trained on visible plus infrared images, on average, are performing better than those trained on the infrared-only and visible-only ones.