Estimating rice yield-related traits using machine learning models integrating hyperspectral and texture features.

Area of Science:

• Agricultural Science
• Plant Phenomics
• Remote Sensing

Background:

• Accurate phenomic diagnosis relies on rapid, precise estimation of multiple trait indicators.
• Developing advanced estimation models for rice yield-related traits (leaf nitrogen concentration - LNC, leaf area index - LAI, aboveground biomass - AGB, grain yield - GY) is crucial.
• A strategy combining spectral data, texture data, dimensionality reduction, and machine learning is key to enhancing estimation accuracy.

Purpose of the Study:

• To enhance the accuracy of estimation models for key rice yield-related trait indicators.
• To investigate the effectiveness of combining spectral and texture data with dimensionality reduction and machine learning techniques.
• To provide a robust method for precise diagnosis of rice traits.

Main Methods:

• Hyperspectral canopy images and trait data (LNC, LAI, AGB, GY) were collected synchronously between 2022-2023.
• Dimensionality reduction techniques (Pearson correlation, SPA, CARS) were applied to select sensitive wavelengths.
• Estimation models were built using ANNs, SVM, 1D-CNN, and LSTM, incorporating spectral and texture features.

Main Results:

• The SPA-ANN model showed the best prediction for LNC (R²=0.82) and LAI (R²=0.75).
• The CARS-ANN model was optimal for AGB (R²=0.90) and GY (R²=0.63).
• Incorporating texture features improved R² by up to 9.9% and reduced RMSE by up to 27.2%.

Conclusions:

• The optimized "spectral + texture + dimensionality reduction + machine learning" method significantly enhances the accuracy of rice trait estimation models.
• This approach provides a valuable scientific basis and technical data for precise diagnosis of rice yield-related traits.