Forest disturbance detection and spatio-temporal dynamic analysis using the GRU-LPETransformer model

Forest disturbance detection is essential for monitoring ecosystem dynamics, as it provides critical insights into the impacts of climate change, natural disasters, and human activities on forest structure and function. Characterizing the spatiotemporal patterns of disturbances is key to assessing carbon sink loss, biodiversity decline, and ecological recovery potential, thereby offering scientific evidence to support global sustainability goals. In this study, we propose the GRU-LPETransformer model, which integrates local temporal feature extraction using a Gated Recurrent Unit (GRU) with global dependency modeling via a Transformer enhanced by learnable positional encoding (LPE). Using Landsat-derived Normalized Burn Ratio (NBR) time series, we conducted forest disturbance detection and spatiotemporal analysis. Training and validation were based on a reference dataset constructed through visual interpretation. Experimental results show that the model achieves an accuracy of 82.59% and an F1-score of 82.96% in binary disturbance classification, outperforming LandTrendr, GRU, and Transformer baselines. For disturbance-year detection, the accuracy within a ±1-year tolerance reaches 74.14%. From 2003 to 2022, forest disturbance in Fujian Province exhibited an overall declining trend, with hotspots concentrated in mid-elevation, moderate-slope areas, reflecting the combined effects of terrain accessibility and human activities. The proposed model offers an efficient and robust tool for forest disturbance detection. In addition to supporting ecological early warning, disaster response, and carbon sink management, it holds promise for application in other ecologically vulnerable regions, thereby contributing to sustainable forest management and decision-making.