An integrated framework for monitoring forest recovery process after disturbance based on Landsat time series images and deep learning methods

The increasing frequency and intensity of forest disturbances, coupled with varying recovery capacities of forest ecosystems, underscore the importance of understanding the full spectrum of disturbance and subsequent recovery dynamics. Previous research has primarily focused on binary forest disturbance mapping, with limited attention given to dynamic disturbance-recovery processes. In this study, we generated a synthetic disturbance-recovery dataset by analyzing the trajectory characteristics of LandTrendr-fitted Normalized Burn Ratio (NBR) time series and capturing multitemporal patterns before, during, and after disturbances. This dataset was used to train an InceptionTime classification model, which was subsequently applied to the LandTrendr-fitted time series to classify different disturbance-recovery processes. Finally, three complementary resilience indicators including Disturbance Magnitude (DM), Year-on-Year Average (YrYr), and State Transition Rate (STR) were calculated to assess forest ecosystem resilience. Results showed that the InceptionTime model achieved higher classification accuracy on the synthetic dataset (OA = 96%), and the overall accuracy of the resulting disturbance-recovery classification map was 91% (±3%). The classification maps indicated that 44.8% disturbance area in the study area exhibited incomplete recovery (DR2), 42.9% completely recovered (DR1), while only a small proportion (6.8%) showed no recovery (DN). The DM value gradually declined over time, while the YrYr value exhibited a subtle upward trend, indicating a progressive enhancement in forest ecosystem resilience. Our study provides a framework for characterizing forest disturbance-recovery processes and assessing ecosystem resilience in the absence of in-situ data, thereby offering valuable insights for forest resource management.