Mapping annual 10-meter evergreen and non-evergreen forests in China from 2017 to 2023 with multi-source remote sensing data

Forests can be classified into evergreen forests and non-evergreen forests based on canopy phenology and foliage persistence. Owing to the complexity of China's terrain and the diversity of its climate, ranging from tropical to cold-temperate, mapping China's evergreen forests presents greater challenges than in tropical and cold regions. Here, we developed an automatic framework to estimate annual 10-meter evergreen and non-evergreen forest cover maps in China from 2017 to 2023 with multi-source remote sensing data. Firstly, by combining multiple land use and land cover (LULC) datasets and time-series tree cover products, 200,193 high-quality training samples of evergreen, non-evergreen, and other land covers were automatically extracted from 2017 to 2023 in China. Subsequently, a novel yearly 10-meter Dynamic World Trees frequency (FQ) index, which effectively captures the phenological characteristics of evergreen forests, was developed. It was used together with the other features in the Random Forest model to generate the initial annual 10-meter evergreen and non-evergreen forest cover maps. A spatio-temporal consistency optimization (SCO) procedure was then applied to eliminate outliers caused by remaining clouds or shadows and missing data in the initial results. Moreover, a novel tree species consistency optimization (TSCO) method was proposed to further modify time-series classification errors between evergreen and non-evergreen forest covers. The proposed framework achieved an overall accuracy exceeding 98% in 5-fold cross-validation, over 94% using the GLanCE validation samples, and reaching 96.18% in an independent validation based on 5,000 visual interpretation samples. Post-processing with SCO and TSCO in the proposed method further improved the accuracy of the initial evergreen and non-evergreen forest mapping results. Moreover, compared to two benchmark products, our results achieve better accuracy and provide more spatial details about the spatio-temporal dynamics of evergreen forests. The main contribution of this study is the development of a robust, automated national-scale evergreen and non-evergreen forest cover mapping framework, providing an accurate and openly accessible baseline for national forest monitoring, ecological assessment, and carbon footprint research. The dataset is publicly available at https://doi.org/10.5281/zenodo.17212181 (Zhao et al. Citation2025b).