Based on machine learning algorithms for estimating leaf phosphorus concentration of rice using optimized spectral...

Area of Science:

• Agricultural Science
• Remote Sensing
• Plant Physiology

Background:

• Accurate estimation of leaf phosphorus concentration (LPC) is essential for optimizing fertilization management and crop monitoring in precision agriculture.
• Spectral reflectance properties of plant leaves are influenced by nutrient status, offering potential for non-destructive LPC assessment.
• Machine learning algorithms can process complex spectral data to build predictive models for plant physiological parameters.

Purpose of the Study:

• To evaluate the effectiveness of full-band spectral data, spectral indices (SIs), and wavelet features for predicting rice leaf phosphorus concentration (LPC).
• To compare the performance of various machine learning algorithms in estimating LPC using different spectral feature sets.
• To identify the optimal combination of spectral features and machine learning models for accurate and reliable LPC estimation in rice.

Main Methods:

• Pot experiments were conducted with varying phosphorus (P) treatments and rice cultivars to collect leaf spectral reflectance and LPC data.
• Spectral data were processed to extract full-band (OR), spectral indices (SIs), and continuous wavelet transform (CWT) features.
• Machine learning algorithms, including random forest (RF), were employed to build predictive models for LPC using the extracted spectral features.

Main Results:

• Phosphorus deficiency altered leaf reflectance, increasing it in the visible and decreasing it in the near-infrared regions.
• The random forest (RF) algorithm combined with spectral indices (SIs) and continuous wavelet transform (CWT) features achieved the highest prediction accuracy (R² = 0.73, RMSE = 0.50 mg g⁻¹).
• The RF model using SIs + CWT outperformed linear regression models based on SIs alone, improving LPC prediction by 32%.

Conclusions:

• Combining spectral indices and continuous wavelet transform features with the random forest algorithm provides a robust approach for remotely estimating rice leaf phosphorus concentration.
• This spectral-based approach offers a valuable tool for large-scale monitoring of rice phosphorus status, supporting precision fertilization and crop management.
• The study highlights the potential of advanced spectral feature extraction and machine learning for non-destructive assessment of plant nutrient levels.