Optical time-series change detection

Optical time-series change detection answers the question: has this surface location changed, and when? Rather than comparing two single-date images, the method builds a per-pixel statistical baseline from the historical time-series of optical reflectance, then tests each new observation against that baseline. Change is not flagged from a single anomalous scene but confirmed only when evidence accumulates across multiple dates, reducing false alarms from cloud contamination and sensor noise.

Two algorithm variants have been deployed in near-real-time deforestation and disturbance alert systems. The GLAD-L alert system (Hansen et al. 2016) applies bagged decision trees to Landsat red, near-infrared, and shortwave-infrared temporal metrics, confirming a disturbance when at least 2 of 4 consecutive observations within a 180-day window cross the change threshold.^[1] The GLAD-S2 extension applies decision trees to Sentinel-2 bands 2, 3, 4, 8, 11, and 12 with a 4-tier confidence scheme (single, low, medium, high).^[2] The RADD alert system (Reiche et al. 2021) takes a different approach: it uses Bayesian conditional probability of forest disturbance, fitting a per-pixel Gaussian mixture model as the temporal baseline and reporting low-confidence detections above 85% and high-confidence detections above 97.5% probability on Sentinel-1 C-band 10m imagery.^[3]

The method applies to any surface monitored by a dense optical or SAR time-series and is not limited to forest. Known failure modes include persistent cloud cover that prevents baseline accumulation, seasonal phenological cycles misread as change, and rapid re-vegetation that obscures confirmed disturbances within the confirmation window.