An improved YOLOv7 network achieved precise mapping of Moso bamboo stem density by UAV combined with Sentinel-2 images

• CACFP-YOLOv7 was proposed for individual tree detection in Moso bamboo forest. • A large number of stem density sample sets were generated in the UAV flight area. • UAV images and satellite images were combined to estimate the stem density. Stem density is an important indicator for investigating aboveground biomass (AGB) and productivity of Moso bamboo forests. Traditional field survey methods are costly and difficult to implement. Thus, this study developed a stem density intelligent estimation technology using the synergy of “unmanned aerial vehicle (UAV)-deep learning-satellite remote sensing” in Anji County, Zhejiang Province. This approach mainly included the following three aspects: 1) Construction of a coordinate attention and center feature pyramid YOLOv7 (CACFP-YOLOv7) model. The CACFP-YOLOv7 incorporates a coordinate attention (CA) mechanism in the backbone to enhance feature extraction capabilities and uses a centralized feature pyramid network (CFPNet) to effectively capture both global contextual information and local edge features. Furthermore, the wise intersection over union (Wise-IoU) loss function is adopted, further improving the model’s training performance. 2) Prediction of stem density in the UAV flight area based on CACFP-YOLOv7 and generation of a large number of stem density samples that match Sentinel-2 remote sensing images. 3) Development of three regression models (random forest (RF), artificial neural network (ANN), and multiple linear regression (MLR)) using these samples to estimate Moso bamboo stem density across the entire study area. The results showed that: 1) The CACFP-YOLOv7 network demonstrated obvious advantages in sample-scale intelligent detection of stem density, with an average precision (AP) of 85.0% for the test set and a smaller root mean square error (RMSE) of 112.69 trees per ha. 2) Among the regression models, the RF model demonstrated the highest accuracy, with an R 2 of 0.753 on the test set, while the MLR model had the lowest accuracy, with an R 2 of 0.523. 3) Based on the RF model, the estimated stem density across Anji County was 2,079 trees/ha, corresponding to a total of 191,572,603 trees. The findings demonstrate that the integrated UAV-satellite remote sensing estimation technique enables efficient and accurate assessment of Moso bamboo stem density.