Extraction of deciduous coniferous forest based on Google earth engine (GEE) and Sentinel-2 image

  Objective  In view of fine monitoring and evaluation needs of forest resources, the application potential of multi-temporal and multi-feature Sentinel-2 images in the identification of deciduous coniferous forests was exploratively studied, and a classification model was built according to the phenological characteristics of deciduous coniferous forests to provide method reference for identifying deciduous coniferous forests on a large scale.

  Method  Based on the GEE platform, the difference between deciduous coniferous forests and other forests in different seasons was analyzed by taking Mengjiagang Forest Farm in Heilongjiang Province of northeastern China as the research area. In this study, four seasonal Sentinel-2 images of spring (May 7 and May 27), summer (August 9), and autumn (October 19) in 2020 were used to construct a multi-feature dataset by extracting spectral features, texture features, and topographic features. Feature optimization was carried out according to feature importance scores. Finally, the optimal model for identifying deciduous coniferous forests was established by a random forest classifier to achieve rapid extraction of deciduous coniferous forest in Mengjiagang Forest Farm.

  Result  The experimental results showed that deciduous coniferous forests displayed obvious vegetation spectral features and seasonal characteristics. The multi-temporal image data contained more phenological period information on deciduous coniferous forests, and the images in spring and autumn can enable better differentiation between deciduous coniferous forests and other forests. In addition, near-infrared and short-wave infrared spectral information can greatly help identify deciduous coniferous forests. By the GEE platform and multi-temporal Sentinel-2 images, it is possible to extract vegetation information efficiently and quickly. The overall extraction accuracy and Kappa coefficient of deciduous coniferous forest reached 91.20% and 0.82, respectively.

  Conclusion  The classification model constructed based on the GEE platform and Sentinel-2 image has certain feasibility and applicability for the rapid extraction of deciduous coniferous forest information, and the research results provide a certain reference value for the large-scale extraction of spatial location distribution information of deciduous coniferous forest.