The implementation of Convention on International Trade in Endangered Species of Wild Fauna and Flora (CITES) to combat illegal logging and associated trade necessitates accurate and efficient field screening of wood species. In this study, a total of 10,237 images of 15 Dalbergia and 11 Pterocarpus species were collected from the transverse surfaces of 417 wood specimens. Three deep learning models were then constructed, trained, and tested with these images to discriminate between timber species. The optimal parameters of the deep learning model were analyzed, and the representative wood anatomical features that were activated by the deep learning models were visualized. The results demonstrated that the overall accuracies of the 26-class, 15-class, and 11-class models were 99.3, 93.7, and 88.4%, respectively. It is suggested that at least 100 high-quality images per species with minimum patch sizes of 1000 × 1000 from more than 10 wood specimens were needed to train reliable and applicable deep learning models. The feature visualization indicated that the vessel groupings and axial parenchyma were the main wood anatomical features activated by the deep learning models. The combination of the state-of-the-art deep learning models, parameter configuration, and feature visualization provide a time- and cost-effective tool for the field screening of wood species to support effective CITES designation and implementation.