Identification of standing dead trees in Robinia pseudoacacia plantations across China’s Loess Plateau using multiple deep learning models

Abstract

Drought-induced tree mortality has increasingly expanded worldwide under the influence of climate warming, with China’s Loess Plateau (CLP) emerging as a critical hotspot for such impacts. As one of the most active tree-planting regions globally, the CLP primarily aims to achieve soil and water conservation despite facing challenges such as limited rainfall and frequent extreme drought events. However, accurate identification of standing dead trees (SDTs) within plantations using remote sensing techniques remains underexplored, and the spatial distribution patterns of SDTs across the CLP are poorly understood. Therefore, this study leveraged unmanned aerial vehicle (UAV) remote sensing to capture high-resolution RGB images of Robinia pseudoacacia plantations. These images were then integrated with a comprehensive evaluation of multiple detection algorithms, including Faster R-CNN, EfficientDet, YOLOv4, YOLOv5, YOLOv8, YOLOv9, and a novel model, YOLOv9-ECA. Particularly, the YOLOv9-ECA was developed by incorporating the ECA module into key network layers to enhance channel dependency modeling and improve feature representation for SDTs detection. Its merit lies in adaptively reweighting feature channels, enabling efficient detection in resource-constrained environments. As expected, the YOLOv9-ECA model demonstrated significant advancements, achieving a detection speed of 123.5f/s, a mAP of 97.8%, and an F1 score of 0.97, outperforming other models in both detection efficiency and accuracy. Subsequently, the model was employed to quantify the spatial distribution of SDTs across the CLP by estimating the number of dead trees per unit area. Results revealed an increasing trend in the number of dead trees per unit along decreasing precipitation gradients, emphasizing the vulnerability of Robinia pseudoacacia plantations in drier regions. Additionally, the number of dead trees per unit varied with slope aspect, with sunny slopes exhibiting the highest values and shady slopes the lowest. This study highlights the potential of YOLOv9-ECA as a powerful tool for the efficient detection of SDTs, offering insights for the sustainable management of Robinia pseudoacacia plantations on the CLP and holding potential applicability to similar environments globally.

Keywords

• Plantation
• Tree mortality
• Drone
• Deep learning
• Management