Leaf‐StarNet: A Lightweight Deep Neural Network for Efficient Plant Leaf Classification

ABSTRACT Plant leaf classification is a core task in biodiversity monitoring, ecological studies, and smart agriculture. In practice, it often still depends on field identification by botanists, which can be slow, laborious, and limited by the availability of skilled personnel. To address these issues, we present Leaf‐StarNet, a compact convolutional neural network tailored for accurate and efficient leaf recognition. The network combines three complementary components: a Lightweight Selective Convolution (LS‐Conv) module for localised and efficient feature extraction, a spatial–channel Squeeze‐and‐Excitation (scSE) attention block to refine both spatial and channel information, and a Frequency‐Based Enhancement (FBE) unit to better capture subtle frequency‐domain patterns. We evaluate the model on the LeafSnap dataset and achieve an accuracy of 96.83%, surpassing lightweight baselines such as MobileNetV2, GhostNetV3, and FasterNet‐T1. Ablation experiments confirm the contribution of each module, while Grad‐CAM visualisations illustrate how the network's focus shifts from fine textures in early layers to higher‐level semantic regions in deeper layers. With its small footprint and low computational demand, Leaf‐StarNet is well suited for deployment on mobile or embedded platforms, providing a practical and reliable solution for automated plant leaf classification.