LiteFormer-EffNet: A Lightweight Transfer Learning Framework for Tree Species Identification from Leaf Images

The main contributions of the paper are summarised as
follows:
• A staged transfer learning workflow for tree species
identification from leaf images is developed.
• Several pretrained models are trained and evaluated on the
Tree Species Identification dataset, and EfficientNetV2M
and MobileNetV2 are identified as the strongest backbone
families.
• The selected pretrained backbones are adapted into four
transfer learning variants, and LiteFormer-EffNet is se
lected as the best-performing variant.
• The ablation-selected LiteFormer-EffNet model is then
trained and evaluated on the LeafSnap field dataset,
demonstrating good cross-dataset performance for tree
species identification.