Detecting Forest Disturbance in Northeast China from GLASS LAI Time Series Data Using a Dynamic Model

Large-scale forest disturbance often leads to changes in forest cover and structure, which imposes a great uncertainty in the estimation of the forest carbon cycle and biomass and affects other applications. In northeastern China, the Daxinganling region has abundant forest resources, where the forest coverage is about 30%. The Global LAnd Surface Satellite (GLASS) leaf area index (LAI) time series data provide important information to monitor the possible change of forests. In this study, we developed a new method to detect forest disturbances using GLASS LAI data over the Daxinganling region of Northeast China. As a dynamic model, the season-trend model has a higher sensitivity toward a seasonal change in LAI. Based on the accumulation of multi-year GLASS LAI products from 1997 to 2002, the dynamic model of LAI time series for each pixel is established first. The time-stepping modeling (TSM) process was designed by using the season-trend method, and sequential tests for detecting disturbances from a time series of pixels. Significant changes in the model parameters were captured as disturbance signals. Then, the near-infrared and shortwave-infrared bands of Moderate Resolution Imaging Spectroradiometer (MODIS) surface reflectance are used as auxiliary information to distinguish the types of forest disturbances. Here, the algorithm led to the detection of two different types of disturbances: fire and other (e.g., insect, drought, deforestation). In this study, we took the forest region as the study area, used the 8-day composite GLASS LAI data at 1000-m spatial resolution to identify each pixel as a fire disturbance, other disturbance, or non-disturbance. Validation was performed using reference burned area data derived from Landsat 30 m imagery. Results were also compared with the MCD64 product. The validation results were based on confusion matrices showing the overall accuracy (OA) exceeded 92% for our method and the MCD64 product. Statistical tests identified that TSM’s product accuracy is higher than that of MCD64. This study demonstrated that the TSM algorithm using a season-trend model provides a simple and automated approach to identify and map forest disturbance