Extreme weather conditions seem to occur stronger and more frequently due to climate change. Very expensive technology is used to predict them, which results problematic when they have to be used in less developed countries. An alternative could be to employ digital sensors, such as phone cameras or existing webcam infrastructures, widely distributed in many countries in the world, and to analyse the captured images. Many methods have been proposed for weather detection including also Convolutional Neural Networks (CNNs). The latter have recently become very popular in the field of computer vision due to their excellent performance in image recognition tasks. However, CNNs are characterized by a high number of learnable parameters that need an equally high number of data points to achieve good performance. Since very big dataset are hardly available, techniques that help to overcome this problem, such as transfer learning, can be used. Different are the transfer learning approaches: the fine-tuning approach, i.e. re-training all the network layers, and the freezing layers approach, i.e. re-training just a subset of the network layers. In collaboration with IBM Netherlands Center for Advance Studies, we optimized a ResNet-18 architecture, modifying the architecture depth and applying regularization methods, to perform weather detection of images showing rain or no-rain conditions. The architecture was previously trained on the ImageNet dataset and then, through the fine-tuning approach, we re-trained the network layers using as training data points weather images captured by webcams distributed in The Netherland, Belgium and London. In particular, we collected a dataset composed by 397041 images showing scenarios such as city roads, urban and rural areas. We also adopted the freezing layer approach on an optimized ResNet-18 architecture and made comparison between the two approaches in relation to weather detection task.