Abstract: Objectives: Mangrove forest density is a critical structural parameter reflecting growth characteristics and serves as an essential indicator for ecological monitoring. However, accurate quantification of mangrove density faces significant challenges due to species diversity, complex stand structures, and dynamic intertidal habitats. This study aims to develop an improved deep learning model for precise multi-scale mangrove tree density estimation in Yingluo Port, China, utilizing readily available high-resolution remote sensing data. Methods: First, a multi-scale mangrove density dataset was constructed using Unmanned Aerial Vehicle (UAV) LiDAR data and Gaussian kernel functions. Subsequently, the standard U²-Net architecture was enhanced by introducing a Multi-scale Dilated Residual Block (MDRB) to capture multi-scale contextual features adaptable to mangrove characteristics, and an Efficient Multi-scale Convolutional Attention Decoding module (EMCAD) to improve multi-scale feature fusion capabilities. Finally, a hybrid loss function was adopted to optimize training, ensuring cross-scale density prediction consistency and thereby generating higherquality density maps. Results: The improved U²-Net model demonstrated robust multi-scale estimation performance. On the 0.2 m resolution dataset, key metrics showed rMAE of 45.12%, rRMSE of 37.18%, and R² of 0.83. On the 0.75 m resolution dataset, results showed rMAE of 11.51%, rRMSE of 14.76%, and R² of 0.92, confirming the model's effectiveness and accuracy. The model was successfully applied to Jilin-1 0.75 m high-resolution satellite imagery, generating a detailed mangrove tree distribution map for Yingluo Port. Prediction results indicate higher mangrove density in core growth areas relative to peripheral zones. Yingluo Port contains approximately 2.67 million mature mangrove trees at an average density of 2436.9 trees per hectare, and localized densities reaching 6720 trees per hectare in core areas. Conclusions: This study presents an effective improved U²-Net model for multi-scale mangrove tree counting. By incorporating MDRB, EMCAD, and hybrid loss function, the model significantly enhances feature extraction capabilities, achieves adaptive fusion of multi-scale features, and enables accurate counting across varying spatial scales. The methodology enables efficient generation of large-scale, precise mangrove distribution maps using costeffective high-resolution satellite imagery, advancing deep learning applications in mangrove ecological monitoring. This tool provides vital data support for scientific protection and sustainable management of mangrove resources.