Optimization of Sassafras tzumu leaves color quantification with UAV RGB imaging and Sassafras-net

• UAV RGB imagery model outperforms in leaf detection/counting. • Sassafras detection and leaf counting accuracy were enhanced by YOLOX and CCTrans. • Sassafras-net model significantly reduces MAE and MSE in leaf counting. • YOLOX minimizes false/missed detections effectively. Quantifying the leaf density and coloration of trees is critical for assessing landscape esthetics and photosynthetic efficiency; however, traditional leaf-counting methods are labor-intensive and potentially harmful to trees, making accurate measurements challenging. To address these issues, we present “Sassafras-net,” an advanced model specifically designed to detect and count colored leaves on Sassafras tzumu trees. The methodology consists of two steps. First, we used an improved model termed YOLOX-CBAM to accurately detect and isolate individual trees. This model proved to be more effective than alternatives, such as YOLOX, YOLOv8, YOLOv7, YOLOv5, and Fater-RCNN. Second, the Sassafras-net model, which is based on the CCTrans network, counts the number of colored leaves per tree. Compared with the original CCTrans model of 52.30 and 84.90, the Sassafras-net model achieved significantly lower mean absolute error and mean squared error values of 27.29 and 39.00, respectively. These results confirm the ability of the model to accurately and efficiently quantify colored leaves. To the best of our knowledge, this is the first study to quantify colored leaves in trees. Our method provides forestry researchers with an effective and economical tool for selecting and breeding S. tzumu trees with enhanced color traits. In addition, this study opens new avenues for studying tree traits related to leaf coloration.