Context-aware quantile baselines for disturbance detection in non-stationary short-rotation forest plantations

Disturbances such as pests, diseases, fire, windthrow, and drought are increasingly affecting forests and require near-real-time monitoring from satellite observations. Existing near-real-time methods often detect anomalies relative to historical vegetation index (VI) baselines based on periodic, parametric, or autoregressive models, which can be difficult to apply in short-rotation plantations with strongly non-stationary canopy dynamics. We introduce the FOrest plantation Canopy ALert system (FOCAL), a monitoring framework designed for relatively homogeneous, intensively managed short-rotation plantations that combines pixel-level Quantile Regression Forest baseline prediction with rule-based detection of persistent negative anomalies. For each new Sentinel-2 acquisition, FOCAL first predicts the range of VI values expected under undisturbed conditions from stand age, management variables, environmental conditions, acquisition geometry, and recent VI observations. A disturbance alert is then issued when observed VI values fall below the lower bound of this predicted contextual range for consecutive acquisitions. FOCAL was evaluated using 118 Eucalyptus plantation stands in Brazil affected by confirmed disturbance events, 141 undisturbed reference stands, and a controlled mortality-patch experiment. Using NDVI, FOCAL achieved a mean class accuracy of 0.78 and a binary F1-score of 0.76, outperforming benchmark time-series methods. Detection was particularly strong for fire (100%) and severe water stress (84%), and mortality patches as small as 12 × 12 m were identified. These results demonstrate FOCAL’s potential for disturbance detection in intensively managed Eucalyptus plantations.