Abstract: Chlorophyll content is an important indicator to characterize the growth status of japonica-rice. Hyperspectral remote sensing technology can obtain the chlorophyll content of japonica-rice leaves speedily without loss. This study used the hyperspectral data of japonica-rice leaf blades in Liaozhong Rice Experimental Station of Shenyang Agricultural University from 2015 to 2017, and used three methods including main component analysis method (PCA),typical correlation analysis method (CCA),nuclear typical association analysis method (KCCA) to reduce dimensions for japonica-rice blade hyperspectral information, and selected the better spectral parameters as the input variable of chlorophyll content inversion model. We used support vector machine regression (SVR), neural network (NN), random forest (RF), least-multiplied (PLSR) four machine learning algorithms to establish an inversion model of chlorophyll content japonica-rice leaves. The results showed that KCCA’s lower-dimensional method had better effect on the hyperspectral reduction of japonica rice leaves than that of PCA and CCA. The model of japonica-rice leaf chlorophyll content inversion model established by KCCA-SVR method had the coefficient of R² =0.801, RMSE=1.610, and the japonica-rice chlorophyll content inversion model had the highest accuracy. The model’s good predictive ability provided data support and model reference for inverse research and nutrient diagnosis of chlorophyll content in japonica-rice leaves.

Key words: chlorophyll retrieval, japonica-rice, hyperspectrum, machine learning