关键词: 森林类型, 变化检测, 地表覆盖, 多源遥感, 深度学习

Abstract:

Objective: The existing automatic forest change detection methods can only detect the change areas, and depend on bitemporal image classification to identify change directions. Most of the data sources are only optical satellite images, so that the frequency of change detection is limited by imaging conditions such as cloudy and foggy weather. The effective synergy of multispectral and SAR images based on data availability to shorten the forest change detection cycle and to build end-to-end models for the change detection of land cover and forest types are essential for dynamically monitoring the forest resource. Method: This study proposes a forest classification and change detection network (FCCDNet), which needs two Sentinel images from different times, either Sentinel-1 or Sentinel-2, as the input and gives three raster maps as output, one with land cover types and one with 0?1 for changed-unchanged classification on each pixel. The model includes a parallel Swin Transformer backbone network automatically for extracting multilevel features of the pre-temporal and post-temporal data, a feature fusion module for generating forest change discrepancy features, and a multitasking learning decoder for simultaneously obtaining the change area and land cover type of the pre-temporal and post-temporal. The model was trained and validated in a study area covering 980 km² in Vastra Gotland in Sweden, using images from Sentinel-1 and Sentinel-2 satellites in 2018 and 2023. A dataset was constructed based on multiple reference data sources, including Swedish national forest inventory data and Swedish forest distribution maps, to train the model. This study compared and validated change detection algorithms such as iterative weighted multivariate change detection (IR-MAD), difference in vegetation index (dVIs), Bitemporal Image Transformer (BIT), and random forest classification comparison method, and analyzed the applicability of forest change detection using different combinations of multispectral and SAR data. Result: The proposed FCCDNet model was able to efficiently and accurately detect forest change areas and directions. When using bitemporal Sentinel-2 images, the classification and change detection accuracies for the two periods of multispectral images were 93.26% and 90.56%, respectively, which were significantly higher than the accuracy using IR-MAD (77.68%), dVIs (78.09), and BIT (79.14%). When the paired Sentinel-1 and Sentinel-2 were used as pre- and post- temporal images, changes were also able to be detected but with lower accuracy (65.94%?76.68%). Conclusion: FCCDNet has achieved the end-to-end forest type change detection with high accuracy and can to some extent solve the problem of forest change detection cycle being limited by the lack of high-quality optical image data, and also shows robustness against salt-and-pepper effects on mapping and exhibits sensitivity to changes with smaller amplitude. The FCCDNet model and data processing framework can support dynamic monitoring of forest resources with high automation and fast response.

Key words: forest type, change detection, land cover, multi-source remote sensing, deep learning