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

Rapidly estimating multiple trait indicators simultaneously, nondestructively, and with high precision is an important means of accurate diagnosis in modern phenomics. Increasing the accuracy of estimation models for rice yield-related trait indicators (leaf nitrogen concentration, LNC; leaf area index, LAI; aboveground biomass, AGB; and grain yield, GY) through a strategy of "spectral data + texture data + dimensionality reduction + machine learning" is highly important.

Between 2022 and 2023, hyperspectral canopy images, the LNC, LAI, AGB, and GY were collected synchronously. Then, dimensionality reduction was performed on the preprocessed spectral data using the Pearson correlation coefficient method, the successive projections algorithm (SPA), and competitive adaptive reweighted sampling (CARS) to select sensitive wavelengths. Estimation models were constructed using artificial neural networks (ANNs), support vector machine regression, one-dimensional convolutional neural networks, and long short-term memory networks. By extracting the texture features corresponding to sensitive wavelengths, high-precision estimation models were constructed using a "spectral data + texture data + dimensionality reduction + machine learning" method.

SPA-ANN provided the best prediction for LNC (R² = 0.82, RMSE = 3.68 g/kg) and LAI (R² = 0.75, RMSE = 0.47), while CARS-ANN was optimal for AGB (R² = 0.90, RMSE = 79.05 g/m2) and GY (R² = 0.63, RMSE = 0.59 t/ha). Adding texture features increased R² by up to 9.9% and reduced RMSE by up to 27.2%.

The optimized method can significantly increase the accuracy of estimation models. The results provide a scientific basis and technical data for the precise diagnosis of rice yield-related traits.