Identification of Weeds Based on Hyperspectral Imaging and Machine Learning

Area of Science:

• Agricultural Science
• Remote Sensing
• Machine Learning

Background:

• Weeds pose significant environmental and economic challenges in New Zealand.
• Traditional weed control methods are labor-intensive, costly, and can have adverse environmental impacts.
• Automated weed identification and mapping are crucial for developing sustainable alternatives.

Purpose of the Study:

• To explore the potential of hyperspectral imaging and machine learning for rapid, automated weed discrimination in pastures.
• To compare the effectiveness of different machine learning models for identifying specific weed species.

Main Methods:

• Acquisition and pre-processing of hyperspectral images from four weed species (yellow bristle grass, wind grass, giant buttercup, Californian thistle).
• Training and evaluation of three classification models: partial least squares-discriminant analysis, support vector machine, and Multilayer Perceptron (MLP).
• Utilizing whole plant averaged spectra (Av) and superpixel averaged spectra (Sp) for model training.

Main Results:

• All tested models achieved repeatable weed identification with overall accuracies ranging from 70-100%.
• The Multilayer Perceptron (MLP) model, using superpixel (Sp) averaged spectra, demonstrated the most robust performance with 89.1% accuracy.
• Four key spectral regions were identified as highly informative for weed characterization.

Conclusions:

• Hyperspectral imaging combined with machine learning provides a viable method for automated weed identification in ryegrass/clover pastures.
• The MLP model with superpixel spectral data shows promise for efficient and accurate weed mapping.
• This technology can support the development of eco-friendly and cost-effective weed management strategies.