Work

DFL-YOLO: Density-Guided Frequency Decoupling and Large Selective Kernels improve dense multi-class tree crown detection

Year: 2026

Type: article

Abstract: UAV-based forestry monitoring is challenged by severe target-scale variation, feature ambiguity caused by crown overlap and background interference, and unstable bounding box localization in crowded scenes. To address these coupled errors, we propose DFL-YOLO, a detection framework organized as a progressive process of scale-context construction, density-conditioned feature disambiguation, and residual localization-error optimization. First, structurally pruned Large Selective Kernel (LSK) modules combine cascaded depthwise convolutions with spatial selection to enrich local and long-range context along selected backbone and neck paths. Second, our Density-Guided Frequency Decoupling Branch (DFDB) predicts a Gaussian-supervised density mask from the resulting context-enhanced P4 feature and applies complementary modulation to its high- and low-frequency components, preserving boundary-sensitive responses while suppressing background interference. The modulated P4 feature is then fused with the P3 and P5 paths and passed to the multiscale detection head. Finally, Multi-Strategy Weighted Intersection over Union (MW-IoU) supervises bounding box regression by combining MPDIoU geometric constraints, the non-monotonic focusing mechanism of Wise-IoU, and task-aligned soft weights. On the Mixed NEON dataset constructed from NEON and Roboflow imagery, DFL-YOLO achieved 82.2% mAP@50 and 54.7% mAP@50:95, representing absolute improvements of 2.6% and 4.2% over the YOLOv11l baseline, respectively. After retraining on VisDrone2019, DFL-YOLO achieved 54.5% mAP@50 and 33.7% mAP@50:95, demonstrating its applicability to dense aerial scenes. Overall, DFL-YOLO provides a robust approach to multi-class tree crown detection in complex forest environments.

Source: Frontiers in Forests and Global Change

DOI: 10.3389/ffgc.2026.1964673

Authors: Yuling Liu, Jiong Mu, Binhong Zhou, Junhao Chen, Xinyue Li, Chunyu Chen

Institution: Sichuan Agricultural University

Language: English

Topic: Advanced Neural Network Applications

Subfield: Computer Vision and Pattern Recognition

Field: Computer Science

Domain: Physical Sciences

SDG: Life on land

Open Access status: gold

Estimated APC paid: $2,353