We address the problem of feature importance. Often, when working with classification or regression problems, the results of black-box deep learning techniques are held to scrutiny in an effort to interpret which and to what extent various features affect outcome. We address this issue specifically when the model has a bottleneck which we will be used to infer feature importance. In this paper, we apply this technique to weather data and study which weather features affect traffic most. To this end, we introduce convolutional spatial embedding to convert data with spatial information into spatial images that are suitable for convolutional neural networks. An advantage of our approach is in dealing with input that has highly correlated features, where removing even an important feature will not increase loss.