A new ensemble framework of online algorithms for continuous monitoring and attribution of vegetation loss agents

• Ensemble of three online algorithms achieved an overall accuracy of 80.1% • kNDVI, NBR and NDMI provided complementary information for disturbance attribution. • CCD contributed to undisturbed areas and BFAST Monitor captured disturbance events. • Disturbance magnitude was the key predictor except in wildfire cases. • Continuous monitoring benefits by combining algorithms for abrupt and subtle change. This study introduces an ensemble framework composed exclusively of online time-series algorithms, for vegetation loss agent attribution in diverse forest and seminatural areas. The study focuses on northeastern Greece, a region with exposed to pressures from different vegetation loss agents —wildfire, drought, and anthropogenic activities— which have not been systematically recorded over the last two decades. To address the spectral variability introduced by each agent across vegetation types, the ensemble incorporates three different Landsat-based spectral indices (kNDVI, NBR, and NDMI) and three online monitoring algorithms —BFAST Monitor (BFM), EWMACD, and CCD — to monitor 16-day observations per year (2000–2024). Each algorithm–index pair formed one of nine base learners (BLs). Their change metrics (breakpoint, magnitude, fitting error, trend, intercept) were fed into a CatBoost meta -learner for four-class one-step attribution. BL-specific CatBoost models were also developed for comparison. The ensemble achieved an overall attribution accuracy of 75.6%, a 6.2% gain, while in disturbance detection it reduced overall error to 14.8%, improving error balance by 7.8%, compared to the most accurate BL. Feature importance analysis identified disturbance magnitudes as the most important metrics in vegetation loss agent attribution. Wildfire was primarily supported by BFM and CCD, drought by the combined responses of BFM and EWMACD, anthropogenic disturbances by BFM and kNDVI-based features, and non-disturbance was characterized predominantly by CCD. These results demonstrate that by combining online algorithms sensitive to both abrupt and subtle changes, overall detection error can be balanced, and agent attribution can be improved without offline analysis.