Spectral Preprocessing Combined with Deep Transfer Learning to Evaluate Chlorophyll Content in Cotton Leaves

Area of Science:

• Agricultural Science
• Spectroscopy
• Machine Learning

Background:

• Accurate chlorophyll content assessment is crucial for cotton's nutritional and physiological evaluation.
• Hyperspectral technology and multivariate analysis are common for chlorophyll detection, but models often lack transferability across different cotton varieties or batches.
• Variations in samples and measurement conditions necessitate robust and adaptable chlorophyll estimation methods.

Purpose of the Study:

• To explore the feasibility of using spectral preprocessing combined with deep transfer learning for effective model transfer in chlorophyll content estimation.
• To investigate the performance of a self-designed convolutional neural network (CNN) for chlorophyll detection and model transfer tasks.
• To identify optimal spectral preprocessing techniques for enhancing model transferability.

Main Methods:

• Seven spectral preprocessing methods were evaluated.
• A self-designed convolutional neural network (CNN) was developed for model building and transfer learning via fine-tuning.
• The combination of first-derivative (FD) and standard normal variate transformation (SNV) was identified as the best preprocessing approach.

Main Results:

• The fine-tuned CNN, using FD + SNV preprocessed spectra, significantly outperformed conventional Partial Least Squares (PLS) and Support Vector Machine Regression (SVR) models on the target dataset.
• Even with a smaller target dataset, the fine-tuned CNN demonstrated superior performance compared to conventional models, achieving satisfactory results.
• Ensemble preprocessing coupled with deep transfer learning proved effective for estimating chlorophyll content across different cotton varieties.

Conclusions:

• Ensemble spectral preprocessing combined with deep transfer learning offers a viable and effective strategy for transferring chlorophyll content estimation models between different cotton varieties.
• This approach provides a cost-effective alternative to developing new models for each batch or variety, enhancing the practicality of field-based cotton nutritional status evaluation.
• Deep transfer learning presents a promising avenue for robust and adaptable hyperspectral data analysis in agriculture.