Automatic evergreen forest mapping based on Harmonised Landsat Sentinel-2 and existing broad-category land cover product within Google Earth Engine

Evergreen forests retain green foliage throughout the year and exhibit distinct phenological and biophysical characteristics. However, although many regional and global land cover products are available at 30 m resolution, most adopt coarse classification schemes and do not explicitly distinguish evergreen forests from other forest types. This study proposes a novel Automatic Evergreen Forest Mapping (AEFM) framework to generate annual evergreen forest maps at the city scale using dense Harmonised Landsat Sentinel-2 time-series data on the Google Earth Engine platform. The framework employs an improved K-means clustering approach to automatically extract high-quality evergreen forest training samples from existing broad-category land cover products. A generalised spectral–temporal feature space is then constructed using comprehensive temporal metrics derived from NDVI, EVI, and LSWI. These features are used to train a pixel-level Random Forest classifier for efficient cloud-based evergreen forest mapping. Applied to Beijing, China, the framework achieved an overall accuracy of 92.29%, with both producer’s and user’s accuracies for evergreen forests exceeding 85%. Comparative analyses show that AEFM outperforms existing approaches. Applications in Qingdao and Zhengzhou yielded overall accuracies of about 90%, demonstrating robustness and scalability. Overall, AEFM provides an automated and reliable solution for fine-resolution evergreen forest mapping and long-term monitoring.