Abstract:

Objective: To enhance the accuracy of forest disturbance detection, a multi-source remote sensing fusion approach was employed to obtain spatiotemporally continuous Landsat imagery. A forest disturbance detection method (Landsat-based deep temporal modeling for forest disturbance detection, LTFD) based on Landsat imagery and deep time-series modeling was proposed. The fire and logging disturbances in the Huzhong area of the Greater Khingan Range were used as the study object to assess the application potential of fused multi-source imagery and examine the advantages of deep learning models in long-term disturbance characterization. This study aims to provide technical references for regional forest disturbance monitoring and sustainable forest management. Method: First, a spatiotemporally continuous Landsat monthly composite dataset for the Huzhong area from 2000 to 2022 was constructed. The spatial-spectral fusion was performed with MODIS 500 m and 250 m spectral bands images to generate MODIS 250 m multispectral imagery, which was subsequently integrated with Landsat data through spatiotemporal fusion to reconstruct cloud-contaminated and missing Landsat observations. The resulting dataset comprised 92 images and provided high temporal resolution support for subsequent forest disturbance detection. Second, a Transformer-based model capable of deep temporal feature representation was introduced for disturbance identification. A large-scale training sample set was built using global forest disturbance products as supervision information. To accurately determine the disturbance onset time, the normalized burn ratio (NBR) threshold method was applied at the pixel-by-pixel extraction. Finally, the LTFD method was validated using visually interpreted fire and logging samples from the Huzhong area, with LandTrendr and random forest (RF) selected as benchmark algorithms for comparative performance evaluation. Result: In this study, the detection performance of the LTFD method was compared with LandTrendr and RF from the perspectives of quantitative evaluation and spatial visual analysis. The results demonstrated that the deep temporal modeling strategy employed in LTFD significantly enhanced disturbance detection accuracy on the constructed monthly composite dataset, and also provided more precise and detailed delineation of disturbance extent and morphology in spatial visualization outputs. In addition, quantitative assessments indicated that the monthly composite data exerted markedly different effects on various disturbance detection approaches, which had a negative impact on LandTrendr. The overall accuracy (OA) was 0.852, a Kappa coefficient was 0.498, and a temporal overall accuracy (TOA) was only 0.088, showing the lowest performance among all methods, suggesting that its piecewise linear fitting strategy is insufficient for accurately identifying disturbance timing. In contrast, the monthly composites benefited both RF and LTFD, with LTFD achieving the highest performance (OA = 0.895, Kappa = 0.701, TOA = 0.767). These results indicated that the Transformer architecture had advantages in temporal dependency and nonlinear modeling with its self attention mechanism. Conclusion: This study verifies the feasibility and effectiveness of integrating multi-source remote sensing data fusion with deep learning models to improve forest disturbance detection accuracy. The findings provide a new technical pathway and methodological reference for developing high-precision and intelligent forest disturbance detection systems, and contribute to advancing regional-scale disturbance monitoring towards greater refinement and intelligence.

Key words: forest disturbance, deep learning, multi-source data, long time-series data, data fusion