Abstract: Although tropical rainforests cover only about 6% of the Earth′s land surface, they harbor more than 50% of the world′s known species and store approximately 25% of the terrestrial vegetation carbon stock, playing an irreplaceable role in maintaining global ecological security. It is essential to accurately obtain the spatial distribution of tree species composition within such ecosystems, as it forms the core foundation for assessing ecosystem functions, monitoring biodiversity, and estimating forest carbon stocks. Due to the widespread phenomena of "same species with different spectra" and "different species with the same spectrum" in rainforest environments, traditional single-source remote sensing data often struggle to provide sufficient spectral and textural discriminative information. Furthermore, conventional convolutional neural networks (CNNs) have inherent limitations in deep feature extraction and multi-modal data synergy, resulting in inadequate generalization capabilities when dealing with highly heterogeneous forest backgrounds. These challenges severely constrain the accuracy of fine-scale tree species classification and the reliability of large-scale mapping applications. Taking the Bawangling and Wuzhishan areas of Hainan Tropical Rainforest National Park as the study region, this research proposes a ConvNeXt model integrated with a Convolutional Block Attention Module (CBAM) for tree species classification in tropical rainforests. By adaptively enhancing the spatial and channel discriminability of feature maps through a dual attention mechanism, the model significantly improves the extraction, fusion, and weighting of multi-source features. Additionally, the synergistic fusion of spatial detail information from Google Earth sub-meter imagery with the temporal spectral characteristics of multi-temporal Sentinel-2 imagery further enhances the accuracy of tropical rainforest tree species classification. The results demonstrate that: 1) Multi-temporal Sentinel-2 data can effectively capture the phenological evolution patterns of different forest stands, achieving better classification accuracy than single-temporal Sentinel-2 data. 2) Multi-source data fusion serves as a crucial approach to overcoming the bottleneck in tree species identification. The synergistic strategy combining "high-resolution spatial details with multi-temporal spectra" outperforms all single-source data schemes in classification performance. 3) The proposed ConvNeXt-CBAM model achieves an overall classification accuracy of 97.27% when processing Google Earth sub-meter imagery, multi-temporal Sentinel-2 data, and multi-source heterogeneous data, which substantially outperforms classic models such as the original ConvNeXt, ResNet-18, and DenseNet-121. These results fully demonstrate that the CBAM attention mechanism can significantly enhance the model′s discriminative capability for complex multi-source features by suppressing background noise and strengthening key feature responses. This study provides reliable technical pathways for tree species classification in complex tropical forest environments, as well as effective support for ecological monitoring, carbon accounting, and sustainable management in Hainan Tropical Rainforest National Park.