A method for continuous sub-annual mapping of forest disturbances using optical time series

Forest disturbances have a major impact on ecosystem dynamics both at local and global scales. Accordingly, it is important to acquire objective information about the location, nature and timing of such events to improve the understanding of their impact, update forest management policies and disturbance mitigation strategies. To this date, remotely sensed data have been widely used for the detection of stand replacing disturbances (SRD) such as windthrows and wildfires. In contrast, less effort has been devoted to the detection of non-stand replacing disturbances (NSRD), typically characterized by slower and gradual temporal dynamics. To address this gap, we propose a method for the automated detection of both SRD and NSRD. The proposed method can detect both past and recent disturbances, with a monthly temporal resolution, in a near real-time fashion by processing new images as they are acquired. Differently from existing approaches that handle the time series as a one-dimensional (1D) temporal trajectory, the method analyzes the sequence of images by organizing them in a two-dimensional (2D) grid-like structure. This representation allows us to model both the intra- and inter-annual variations of the time series taking advantage of the annual cyclical nature of the plant phenology. The method has been tested on study areas attacked by bark beetles achieving a user’s accuracy and producer’s accuracy of 0.91±0.08 and 0.81±0.07 (with 95% confidence intervals) for the disturbed areas, respectively.