Temporal Anomaly Detection in Multispectral Imagery

In spectral image analysis, anomaly detection algorithms implicitly perform spatial anomaly detection. That is, the pixels are assumed to all be from the same image at a single time, and the anomalies are pixels whose spectra differ from the rest of the pixel spectra in the scene. In this research we are instead interested in temporal anomalies, where we compare a single pixel against itself over time rather than against the rest of the pixels in the scene. To do this requires a large sample of registered images taken at different points in time, which is easily enabled by global-coverage constellations such as Sentinel-2 and Landsat. In doing so we can implement the same anomaly detection algorithms that are traditionally applied spatially, but draw our “background” distribution from time instead of space. Here we present a straightforward experiment with Temporal RX, where we use the traditional RX algorithm but with a temporal background distribution. We use the annual Burning Man music festival as our test case and explore the spatially-aggregated temporal anomaly detection results.