The automatic recognition of weather in images has many important applications in different fields, such as: land and air traffic control, autonomous vehicles, road safety warnings, crop control, improvement of images taken in outdoor areas, among others. Despite the great applicability, this field of study has not yet been explored in detail, primarily due to the great challenge and difficulty involved in extracting deterministic features for each type of weather. Several works have focused their efforts on designing binary classifiers that allow determining just two classes. A difficulty lies especially in the fact that the target classes are not completely exclusive in an image. Different classes can share the same features. Another difficulty that previous work has faced is the need for a large number of labeled images to model the various weather states. In this work, we propose an approach called self-supervised deep learning applied to weather recognition in order to reduce the requirement of the huge amount of labeled images. Our architecture, a ResNet-50 implementation, is responsible for obtaining the representations of each unlabeled image with a self-supervised approach for both pre-training and fine-tuning steps. It has been used transfer learning for sharing the architecture between these steps. Our results reached an average accuracy of 0.8833. Based on this result, it can be concluded that self-supervised learning is a convenient solution to obtain high performance in the weather recognition task from digital images.