There is a correlation between temporal patterns of geospatial activity and land use type. A novel self-supervised approach has been proposed for landscape surveys based on activity time series, in which the time series signal is transformed into a frequency domain and compressed into weddings by a shrinking autoencoder, which preserves the periodic temporal patterns observed in the time series. Decorations are inputs to a segmentation neural network for binary classification. Experiments show that time weddings are effective in classifying residential and commercial areas.