Machine learning-based Alpine treeline ecotone detection on Xue Mountain in Taiwan

Abstract. Taiwan is characterized by high mountains density, with over 200 peaks exceeding 3000 m in elevation. The alpine treeline ecotone (ATE) is a transitional zone between different vegetation types. The species distribution, range variations, and movement patterns of vegetation within the ATE are crucial indicators for assessing the impact of climate change and warming on alpine ecosystems. Therefore, this study focuses on the Xue Mountain glacial cirques in Taiwan (approximately 400 ha) and utilizes WorldView-2 satellite images from 2012 and 2021 to compute various vegetation indices and texture features (GLCM). By integrating these features with Random Forest (RF) and U-Net models, we developed a classification map of the ATE in Xue Mountain. We analyzed changes in bare land, forest, krummholz, and shadows within the ATE from 2012 to 2021. The results indicate that the classification accuracy reached an overall accuracy (OA) of 0.838 when incorporating raw spectral bands along with vegetation indices and texture features (GLCM) (77 features in total). Feature importance ranking and selection reduced training time by 14.3 % while ensuring alignment between field survey treeline positions and classification results. From 2012 to 2021, tree cover density increased, with the total forest area expanding by approximately 10.09 ha. The upper limit of forest distribution shifted upslope by 32.00 ± 4.00 m, with the most significant area changes occurring between 3500 and 3600 m, while the 3700 to 3800 m range remained relatively stable. This study integrates remote sensing imagery with deep learning classification methods to establish a large-scale ATE classification map. The findings provide a valuable reference for the sustainable management of alpine ecosystems in the Xue Mountain glacial cirques in Taiwan.