An improved hybrid approach involving deep learning for urban greening tree species classification with Pléiades Neo 4 imagery—A case study from Nanjing, Eastern China

作者：Min Sun, Stephane G.P. Debulois, Zhengnan Zhang, Xiaolei Cui, Zhili Chen, Mingshi Li · 发表于：Ecological Informatics · 年份：2025 · DOI：10.1016/j.ecoinf.2025.103331 · 被引用次数：2 · 研究领域：Remote Sensing in Agriculture、Land Use and Ecosystem Services、Remote-Sensing Image Classification

The distribution of urban tree species shapes spatial differences in urban eco-environmental benefits and liveability. Although very high-resolution (VHR) multispectral optical satellite imagery has significant potential for detailed urban tree species classification, few studies have reliably classified urban tree species in shadow areas. This study proposed an improved hybrid approach for the accurate classification of target urban tree species using Pléiades Neo 4 imagery. We compared deep learning (DL)-based shadow detection methods with classical index methods in urban green spaces to identify shaded vegetation areas, with the optimal method being selected for subsequent use. We developed an ensemble model consisting of machine learning (ML) and DL approaches with two enhancement strategies, namely an attention mechanism and a fixed category weighting scheme (i.e., a weighted dictionary), which reweights classifier outputs based on the classification accuracy of each category. We then evaluated the effectiveness of the ensemble model by comparing it with four base models, using cluster-level feature datasets derived from superpixel segmentation for tree species classification in shadow areas. Tree species in sunlit areas were classified using the UNet 3+ model, and the results were spatially integrated to determine the overall tree species distribution. Our ensemble model achieved an overall accuracy (OA) of 66.19 %, outperforming four widely used base models (OA = 56.67...