Evaluating landsat time series of above-ground biomass to monitor boreal forest recovery following different types of disturbance

Global changes are exerting uncertain effects on boreal forest dynamics, highlighting the need for accurate methods to monitor above-ground biomass (AGB). Numerous studies have developed AGB estimation models based on Landsat reflectance and then applied these models across time series. However, only few studies provide an accuracy assessment of the resulting AGB time series using field plot data, limiting the ability to evaluate their applicability and highlight potential constraints. We developed AGB maps using a two-phase approach combining wall-to-wall airborne LiDAR point cloud data, winter and summer Landsat composites time series and Unet convolutional neural networks. This 36 years AGB time series (1985–2020) was evaluated using over 3,000 permanent sample plots with periodic measurement. The accuracy assessment of the produced maps indicates an R2 of 0.50 and RMSE of 33 t/ha. Analyses of dynamics after severe disturbance (harvesting and fires) reveal a close similarity in trajectories between permanent sample plots and predicted AGB values (Pearson correlation of 0.93). The maps enabled identifying stands where recovery failed, highlighting potential of this approach for monitoring disturbance recovery. When comparing our method to NBR spectral recovery, we found that Unet-based AGB provided a more realistic approach for monitoring structural recovery.