A hyperspectral inversion model for predicting chlorophyll content of Longan leaves based on deep learning

Objective  To study the distribution of chlorophyll content of Longan (Dimocarpus longan Lour) leaves in different growth periods, realize non-destructive measurement of the influence of pests and diseases on chlorophyll distribution, and provide a reference for evaluating the cold-resistant ability of young leaves, fertilizing amount in the fruiting period and pruning of mature leaves.

Method  Hyperspectral images of Longan leaves in three growth periods were acquired via an online hyperspectral imaging system within the spectral region of 369–988 nm wavelength. An automatic masking method was used to extract the interest regions. The chlorophyll content was measured by the spectrophotometric method. The relationships between the spectral response characteristics and chlorophyll contents of Longan leaves in three growth periods were measured based on Pearson correlation coefficient (r). A partial least squares regression (LSR) model was established. The relationship between the texture feature of selected image and chlorophyll content was analyzed. The spectroscopy and texture features were imported to the spare auto-encoder (SAE) model in deep learning to predict the chlorophyll content of Longan leaves. The distribution of chlorophyll content was predicted using SAE model based on the mapping information.

Result  The peaks of correlation coefficient curves of Longan leaves in three growth periods appeared in the vicinity of 700 nm. The wavelength of the highest correlation coefficient for young, mature and old ripe leaves was 692, 698 and 705 nm, respectively. The correlation coefficient (r) of the most sensitive band in full period was higher than those in three growth periods, which was up to 0.890 3. Among all regression models, the prediction effect of LSR model based on the absorption band of the minimum reflectivity and total reflectivity was the best (R²_c=0.856 8, RMSEc=0.219 5; R²_v=0.771 2, RMSEv=0.286 2), and the determination coefficients of its calibration and validation sets were higher than those based on a single parameter. SAE model importing spectroscopy and texture features performed the best (R²_c=0.979 6, RMSEc=0.171 2; R²_v=0.911 2, RMSEv=0.211 5) and the most stable to predict chlorophyll contents of Longan leaves in different growth periods, its standard deviation was only 29.9% of LSR model.

Conclusion  A method automatically extracting interest region was proposed, its success rate was 100%. The performance of SAE model based on spectroscopy and texture features was more stable than those of regression models based on spectroscopy to predict chlorophyll contents of Longan leaves in different growth periods. SAE model is suitable for predicting the distribution of chlorophyll content of Longan leaves as a non-destructive method.