Advancing Wheat Variety Identification: CSKNN Leveraging Hyperspectral Imaging

By Dr. Sampath LonkaReviewed by Susha Cheriyedath, M.Sc.Aug 22 2023

Hyperspectral imaging plays a crucial role in rapidly identifying different wheat varieties. However, challenges such as noise and limited spatial utilization in existing methods have hindered progress in this area. In a recent publication in the journal Computers and Electrical Engineering, researchers introduced a method called Cost-sensitive K-Nearest Neighbor using Hyperspectral Imaging (CSKNN) for identifying wheat seed varieties.

Background

Wheat, a vital global food crop, particularly in northern China, faces issues related to seed mixing due to its diverse varieties. Targeted storage and processing are essential for different wheat types. Hyperspectral imaging holds promise for identifying wheat varieties, yet challenges persist.

Complex seed structures, spectral variation, and traditional identification methods' limitations hinder robust modeling. Machine learning and deep learning techniques are emerging for wheat seed identification, but manual feature extraction in machine learning and data scarcity in deep learning pose challenges. The algorithm K-nearest neighbor (KNN) offers a simple solution. Enhanced KNN methods tackle memory consumption and class overlap, while ensemble-based KNN addresses incremental learning on data streams with new classes. Despite these advancements, challenges in accurately identifying wheat varieties persist, warranting further research.

Wheat variety identification using CSKNN

The current study introduces the CSKNN method for identifying wheat seed varieties using hyperspectral imaging. The CSKNN process for wheat variety identification involves three principal stages: (1) data smoothing and denoising (SDOD), (2) valuable feature extraction (EOVF), and (3) cost-sensitive KNN (CSKNN).

SDOD encompasses region of interest cropping followed by spatial denoising to mitigate noise and enhance signal quality. EOVF focuses on feature extraction, employing supervised learning linear discriminant analysis (LDA) to distill informative features from the denoised data, thereby aiding wheat seed identification. Subsequently, CSKNN is deployed for the final classification of wheat varieties.

In the SDOD phase, the emphasis is on refining the raw spectral data through smoothing and denoising, a pivotal process for subsequent feature extraction. Conventional filtering techniques fall short of harnessing spatial information in hyperspectral imagery. To effectively leverage spatial information while minimizing noise, a spatial denoising strategy (SDS) is introduced. SDS adopts a polynomial fitting approach grounded in the least squares method, facilitating spatial filtering by employing convolution operations based on a two-dimensional fast Fourier transform.

The EOVF phase capitalizes on the refined data by emphasizing valuable feature extraction. Supervised learning LDA, a linear dimensionality reduction method, is employed to enhance data analysis accuracy and circumvent the curse of dimensionality. LDA maximizes inter-class variance while minimizing intra-class variance post-projection, effectively improving the distinguishability of samples.

The final phase, CSKNN, aims to enhance the capacity of traditional KNN's capacity by accounting for class imbalances and the associated misclassification costs. A distance-based cost function is introduced, underpinned by a logarithmic transformation, to adapt the classifier based on the principle of minimum expected cost. It incorporates factors such as distance influence and cost weight, promoting a more robust classification, particularly for minority-class samples.

Influence of feature extraction and k value on CSKNN

The dataset comprises 10 high-quality wheat varieties from Henan. Utilizing the portable visible and near-infrared imaging spectrometer system Surface Optics Corporation (SOC) 710, featuring a spectral range of 400–1000 nm and 128 bands, data collection encompassed three strains: Zhoumai (Zhoumai 28, 36, 27), Zhengmai (Zhengmai 101, 7698, 366), and Bainong (Bainong 207, 4199, 307, 58). Each variety comprises 120 samples, with 128 bands per sample. Representative bands were chosen between one and 128.

To facilitate subsequent classification, the wheat spectral information is extracted from the background. Spectra after region-of-interest extraction display a consistent trend across samples, with a gradual rise between 400 nm and 650 nm and a distinct absorption peak within the 650 nm to 700 nm range. The interval of 700 nm to 1000 nm displays a gradual decline. Distinct spectral profiles for different varieties arise from compositional differences, enabling wheat variety differentiation.

An Investigation into the influence of extracted feature count and k size on CSKNN's classification performance is undertaken. Results indicate that as the training sample number increases, CSKNN performance improves. Classification performance increases initially with feature count, stabilizing eventually. Optimal feature compression is achieved with eight features. Similarly, the effect of the value of k on classification accuracy is examined. Optimal performance is observed with a k value of seven, outweighing the significance of the feature extraction count.

CSKNN's classification performance is compared against several machine learning and deep learning methods for both different strains and varieties. CSKNN consistently outperforms most methods, displaying superior classification ability. Classification performance slightly diminishes for different varieties, but CSKNN remains superior to competing methods, particularly for wheat species identification.

Algorithm stability is assessed using mean square error. Deep learning methods exhibit better stability compared to traditional approaches. However, CSKNN, leveraging cost sensitivity, maintains superior algorithmic stability even in comparison with deep learning methods.

An ablation study is conducted on 10 wheat variety datasets to showcase the impact of key steps on CSKNN. Each step, including data smoothing and denoising, feature extraction, and cost sensitivity, contributes positively to classification performance enhancement. The full CSKNN model, encompassing all steps, demonstrates the best classification performance, validating the positive impact of each key step.

Conclusion

In summary, researchers introduced a CSKNN approach for wheat seed identification using hyperspectral imaging. A noise elimination strategy is employed, followed by feature compression via LDA. Cost-sensitivity enhances traditional KNN, improving classification. Experiments demonstrate CSKNN's strong discrimination against wheat strains and varieties in the same area. Refinement of the model is needed for identifying multiple wheat varieties in one region. Data validation from diverse regions is absent but planned for future research.