*=Equal Contributors
There exists a correlation between geospatial exercise temporal patterns and kind of land use. A novel self-supervised strategy is proposed to survey panorama primarily based on exercise time collection, the place time collection sign is remodeled to frequency area and compressed into embeddings by a contractive autoencoder, which protect cyclic temporal patterns noticed in time collection. The embeddings are enter to segmentation neural community for binary classification. Experiments present that the temporal embeddings are efficient in classifying residential space and industrial space.
