Urban forest tree species classification and change monitoring in Xi’an based on multi-source time-series features

Abstract

Urban forests are critical infrastructure within urban ecosystems and provide multiple ecosystem services, including heat island mitigation and enhanced carbon storage. However, classification accuracy and transferability in cities are constrained by high landscape heterogeneity, patch fragmentation, mixed pixels, and spectral confusion among different land-cover types. To improve forest-classification robustness under highly heterogeneous urban conditions, this study used Xi’an, China, as the study area and constructed a 2020–2025 multi-source remote sensing time-series dataset on the GEE platform. Ninety features were extracted from seasonal median composites, and collinear variables were removed using the variance inflation factor (VIF). The study compared three models, namely RF, GBT, and CART, selected the best-performing RF model, and further optimized it using recursive feature elimination (RFE), ultimately identifying 58 key features. The results show that: (1) RF (RFE) achieved the highest accuracy (OA = 77.40%, Kappa = 0.7175), and the original feature set was reduced from 90 dimensions to an optimal 58-dimensional feature subset; (2) topographic factors (elevation and slope) and third-quarter vegetation indices (GNDVI, NDVI, and ARVI) were the core features; (3) the spatial pattern of forests followed a ‘southern forests, northern grasslands; denser in the south, sparser in the north’ configuration, with evergreen needleleaf forest concentrated in the southern mountains, other vegetation occurring mostly as patchy distributions across the northern plain and peri-urban areas, and non-vegetated areas concentrated in the central-northern core; and (4) From 2020 to 2025, evergreen needleleaf forest first decreased and then increased; deciduous broadleaf forest fluctuated at a high level; evergreen broadleaf forest continued to shrink; other vegetation remained generally stable; and all land-cover classes exhibited differentiated temporal trajectories. The automated mapping framework based on the GEE machine-learning model developed in this study can provide quantitative and continuously updatable data support for urban forest planning and ecological performance assessment in Xi’an.