Plant Species Classification Based on Hyperspectral Imaging via a Lightweight Convolutional Neural Network Model.

Area of Science:

• Botany
• Computer Science
• Remote Sensing

Background:

• Traditional RGB imaging struggles with plant species classification due to limited spectral information.
• Existing methods often focus on single leaves, neglecting live-crown features and complex spectral patterns.
• Similar-looking species pose challenges for RGB-based classification algorithms.

Purpose of the Study:

• To develop a novel framework for plant image classification using hyperspectral imaging (HSI) and deep learning.
• To evaluate the performance of a lightweight conventional neural network (LtCNN) for plant species identification.
• To investigate the impact of different spectral band combinations on classification accuracy.

Main Methods:

• A plant image dataset of 1,500 images across 30 species was created using a 470-900 nm hyperspectral camera.
• A lightweight conventional neural network (LtCNN) model was designed for image classification.
• Comparative analysis with state-of-the-art CNNs and evaluation of various band combinations were performed.

Main Results:

• Hyperspectral imaging achieved higher accuracy than simulated RGB images (kappa=0.95 vs. kappa=0.90).
• Combining RGB and near-infrared bands improved classification to kappa=0.95.
• The proposed LtCNN model achieved a kappa of 0.95 using critical spectral features (green-edge, red-edge, near-infrared bands).

Conclusions:

• Hyperspectral imaging and deep learning offer a robust solution for plant species classification.
• The LtCNN model demonstrates high adaptability for live-crown image analysis with fewer training samples.
• This approach overcomes limitations of RGB imaging and traditional feature engineering for plant identification.