Python package for near real time detection of change in spatio-temporal datasets

nrt provides a standardized interface for Near Real Time monitoring of disturbances on satellite image time-series. The package is optimized for fast computation and suitable for operational deployment at scale. Five monitoring frameworks from scientific literature on change detection are implemented and exposed via a common API. All five monitoring framework share a common general approach which consists in modelling the "normal" behavior of the variable through time by fitting a linear model on a user defined stable history period and monitoring until a "break" is detected. Monitoring starts right after the stable history period, and for each new incoming observation the observed value is compared to the predicted "normal" behavior. When observations and predictions diverge, a "break" is detected. A confirmed "break" typically requires several successive diverging observations, this sensitivity or rapid detection capacity depending on many variables such as the algorithm, its fitting and monitoring parameters, the noise level of the history period or the magnitude of the divergence. The five monitoring frameworks implemented are:

• Exponentially Weighted Moving Average (EWMA) (Brooks et al., 2013)
• Cumulative Sum of Residual (CuSum) (Verbesselt et al., 2012; Zeileis et al., 2005). CuSum is one of the monitoring option of the bfastmonitor function available in the R package bfast.
• Moving Sum of Residuals (MoSum) (Verbesselt et al., 2012; Zeileis et al., 2005). MoSum is one of the monitoring option of the bfastmonitor function available in the R package bfast.
• Continuous Change Detection and Classification of land cover (CCDC, CMFDA) (Zhu et al., 2012, 2014) - Partial implementation only of the original published method.
• InterQuantile Range (IQR) - Simple, unpublished outlier identification strategy described on stackexchange.