Adapting pre-trained large model to robust individual tree crown detection from UAV optical images

Remarkable success has been demonstrated by pre-trained foundation models across a variety of various downstream vision problems. To address the intensive resource demands of full-parameter fine-tuning, these models are adapted by Parameter-Efficient Fine-Tuning (PEFT) methods, wherein only a small subset of parameters is updated. However, the application and effectiveness of PEFT methods in specialized domains like forestry monitoring–particularly for tree crown identification at the individual level in UAV imagery–have not been extensively investigated. In this study, TreeAdapter, a novel and efficient PEFT method, is introduced for the adaptation of large vision-language models for tree crown detection. The approach is composed of two key modules. First, Frequency-Enhanced Visual Prompting (FEVP) is proposed, a module in which visual prompts are constructed from high-frequency features to explicitly strengthen the model’s awareness of fine-grained crown edge information. Second, the Perception-Enhanced Adapter (PEA) is introduced, through which multi-scale depthwise convolutions are integrated into the Swin Transformer backbone to augment its ability to capture local textural and structural details. The proposed method is validated through extensive cross-domain experiments conducted on four datasets representing diverse forest structures. Promising zero-shot cross-domain transferability is achieved by TreeAdapter, as demonstrated by results that outperform existing deep learning and PEFT methods. Through these findings, TreeAdapter is established as an effective PEFT approach, capable of efficiently adapting the powerful capabilities of foundation models to the unique task in forestry remote sensing monitoring. Our code is released at https://github.com/jiuyuzhang/Tree-Adapter .