Assessing different regression algorithms for paddy rice leaf nitrogen concentration estimations from the...

Area of Science:

• Agricultural Science
• Remote Sensing
• Spectroscopy

Background:

• Non-destructive estimation of crop leaf nitrogen concentration (LNC) is crucial for quality assessment and precise nitrogen (N) fertilizer management.
• First derivative analysis effectively reduces spectral noise, enhancing the estimation of leaf N and chlorophyll concentrations under varying fertilization levels.

Purpose of the Study:

• To evaluate the effectiveness of various regression algorithms combined with first-derivative fluorescence spectra (FDFS) for estimating LNC in paddy rice.
• To optimize the parameters of these regression models and compare their performance for LNC estimation.

Main Methods:

• Laser-induced fluorescence (LIF) spectra were processed to obtain first-derivative fluorescence spectra (FDFS).
• FDFS data were used as input for regression models including Principal Component Analysis (PCA), Partial Least-Square Regression (PLSR), Random Forest (RF), Radial Basic Function Neural Network (RBF-NN), and Back-Propagation Neural Network (BPNN).
• Model parameters were optimized, and the performance of different models (e.g., PCA-RBFNN, PLSR) was compared based on R-squared (R²) and standard deviation (SD) values.

Main Results:

• Principal Component Analysis (PCA) effectively extracted key spectral information, improving the stability and robustness of regression models.
• The PCA-RBFNN model demonstrated superior performance for LNC estimation, achieving the highest average R² (0.8743) and lowest SD (0.0256).
• Partial Least-Square Regression (PLSR) also showed promising results, with an average R² of 0.8412, performing better than other models except PCA-RBFNN.

Conclusions:

• PCA preprocessing enhances the stability and accuracy of spectral analysis for LNC estimation.
• The PCA-RBFNN model shows significant potential for accurate and non-destructive LNC estimation in paddy rice.
• FDFS combined with optimized regression models provides a robust approach for precision agriculture applications.