Abstract
Abstract
Since 2023, although Machine Learning (ML) has been widely applied to vegetation extraction from Remote Sensing Images (RSI), most research has focused on static coverage. Despite recent progress in time-series monitoring and multi-technology integration, persistent issues remain, including low accuracy in quantifying recovery trajectories, inadequate characterization of temporal evolution, and poor adaptability in multi-source data fusion. These limitations hinder precise habitat monitoring along transmission corridors. To address these challenges, this study develops an adaptive recovery trajectory modeling technique for multi-source disturbance fusion. By integrating RSI and LiDAR data, a dynamic fusion method for multi-source remote sensing time-series features is proposed, and a recovery trajectory model that incorporates quantitative construction disturbance factors is constructed. A multi-scale temporal attention mechanism is designed to enhance the detection of key turning points. Additionally, a corridor-specific application system ensures alignment with monitoring scenarios. The core innovations include adaptive stage-wise weight adjustment for multi-source time-series fusion and the introduction of disturbance adaptation factors, which improve trajectory description accuracy. Simulation experiments show that the proposed technique achieves a comprehensive performance score of 0.889–0.905, outperforming YOLOv10 and ConvLSTM by 13.7–23.5 %. Key recovery trajectory turning points are identified with 91.5 % accuracy, the mean absolute error (MAE) is reduced by 32.1–49.4 %, and processing time (RT) is shortened by 26.0 % compared to LiDAR + Optical RSI fusion. Robustness reaches 0.915, and adaptability across three typical monitoring scenes averages 0.889. This technique fills a gap in quantitative habitat recovery modeling for transmission corridors, supports dynamic vegetation monitoring and regulation, and advances intelligent ecological maintenance in power grid operations.