A Deep Learning Framework for Processing and Classification of Hyperspectral Rice Seed Images Grown under High Day...

Area of Science:

• Agricultural Science
• Computer Science
• Image Processing

Background:

• Hyperspectral imaging (HSI) offers rich spectral information for seed analysis.
• Deep learning models can effectively process complex HSI data for classification tasks.
• Assessing rice seed response to heat stress is crucial for crop improvement.

Purpose of the Study:

• To develop and validate a framework combining HSI and deep learning for rice seed classification.
• To evaluate seed-based and pixel-based deep learning approaches under various heat stress conditions.
• To create a user-friendly software application for HSI-based seed analysis.

Main Methods:

• A seed-based approach utilizing a 3D convolutional neural network (3D-CNN) on full spectral hypercubes.
• A pixel-based approach employing a deep neural network (DNN).
• Validation using hyperspectral images of rice seeds subjected to controlled high-temperature durations (day, night, both).

Main Results:

• The 3D-CNN achieved average accuracies of 91.33% and 89.50% for treatments and durations, respectively.
• The DNN achieved average accuracies of 94.83% and 91% for treatments and durations, respectively.
• Both models demonstrated superior performance compared to existing literature.

Conclusions:

• The integrated HSI and deep learning framework effectively classifies rice seeds under heat stress.
• The developed deep learning models offer high accuracy and potential for broader application in seed research.
• This approach can be extended to study temperature tolerance in other rice cultivars.