3-D–2-D Hybrid Lightweight CNN Model: Enhancing Canopy Feature Retrieval in Hyperspectral Imaging for Accurate Plant Species Classification

Plant species recognition is essential for effective resource and environmental management, making it a key area of research in remote sensing. Deep learning (DL), particularly Convolutional Neural Networks (CNNs), has been widely used to identify images of plant organs and canopies from various sensor-derived images. However, the scalability and cost-effectiveness of CNNs in recognizing a large number of species remain underexplored. Key challenges include performance limitations when handling many species, the impact of feature similarity on identification, overfitting and scalability issues, and the trade-off between data cost and performance. To address these challenges, we constructed a comprehensive hyperspectral dataset comprising 100 subtropical plant species with consistent crown-level imaging. We propose a novel 3D–2D hybrid lightweight CNN model, named Hybrid-LtCNN, designed to jointly capture spectral–spatial and semantic texture information while reducing overfitting and computational burden. The proposed model was compared against five existing DL models (GoogleNet, AlexNet, MobileNetV3-Small, LtCNN, and TinyViT). Experimental results show that Hybrid-LtCNN achieved an impressive macro F1-score of 0.9935 without signs of overfitting, significantly outperforming the baseline models. Furthermore, Class Activation Mapping (CAM)-based visualizations indicate that the model successfully captured crucial plant-specific features aligned with traditional taxonomy. We also investigated the impact of spectral band selection on classification performance, as well as the sensitivity of various CNN models to datasets with different numbers of species. These findings highlight Hybrid-LtCNN's scalability, interpretability, and potential for practical applications in remote sensing-based plant species classification.