柑橘病害近红外光谱判别模型及病害对理化指标影响

柑橘病害严重影响柑橘产量与品质，为进一步寻找叶片光谱与其理化指标变化规律，故将近红外光谱技术与化学计量学相结合探索柑橘病害光谱判别及病害对理化指标影响的可行性。利用便携式近红外光谱仪，获取柑橘正常、溃疡病和砂皮病3类叶片光谱并测量理化指标值，开展光谱特性与理化指标规律分析，进行SPA(successive projections algorithm)与PCA（principal component analysis）变量筛选，结合RF （random forest）与LWPLS（locally weighted partial least squares）分别建立柑橘病害定性模型及理化指标定量模型。对比分析模型结果，发现基于原始光谱变量的LWPLS定性模型效果最佳，其判别准确率为94.03%。用401个光谱变量为输入，基于LWPLS的正常叶片SPAD值模型分析结果最优；虽正常、溃疡病和砂皮病3类叶片综合的LWPLS定量模型RMSEP较大为4.46%，但模型对叶片SPAD值具有较好的预测精度，R²和RPD分别为0.93、2.19。研究表明，近红外光谱技术结合化学计量学判别柑橘病害及分析病害对叶片理化指标的影响具有一定可行性，可为柑橘病害实时现场检测提供重要科学参考。

Model for identification of citrus diseases by near infrared spectroscopy and influence of citrus diseases on physicochemical index

Pest and diseases have seriously threatened to the yield and quality of citrus. This study aims to explore the effect of citrus diseases on SPAD value using identification model by near-infrared spectroscopy with chemometrics. Leaf spectra and physicochemical indexes were further determined to detect the citrus diseases. The NIR spectra of citrus leaves with normal, canker and emery derma were obtained to measure the physicochemical indices by portable NIR spectrometer. The spectrum analysis showed that there was a reflection peak in the leaves near 553 nm, indicating the reflection peak after the chlorophyll saturated the high-frequency absorption of light. The maximum reflection peak of normal leaves was about 0.21, whereas, the lowest was about 0.13 in the leaves with the sand skin disease. The reflectance of normal leaves was higher than that of diseased leaves at 750-900 nm. The reason was that the reflectance decreased to enlarge the leaf cell cavities under disease stress. The variance was analyzed for two physical and chemical indexes. There were significant differences in the SPAD value indexes of the three types of citrus leaves, while no significant difference was found in fresh quality indexes. A qualitative and quantitative model of citrus diseases was also established to clarify the relationship with leaf spectrum. According to the ratio of 2:1, 212 blades were randomly divided into two dataset with 145 samples in training set and 67 samples in verification set. The SPA and PCA were used to screen the spectral variables. 17 and 20 optimal variables were selected to combine the RF and LWPLS. The comparison showed that the best detection was achieved in the qualitative model of LWPLS for three types of disease leaves using 401 variables of the original spectrum, with the accuracy of 94.03%. While the low accuracy of 77.61% was found in the qualitative model of RF for three types of disease leaves using SPA screening 17 variables. The main reason was that a few sensitive variables were eliminated, although the characteristic variables represented the most information of citrus leaves. A numerical model was also constructed to further verify whether the leaf diseases affected the SPAD value of normal and diseased leaves. The best performance was achieved in the LWPLS model in the normal blade SPAD value analysis with 401 spectral variables as the input. The RMSEP of LWPLS quantitative model for normal, canker and emery was 4.46%, the model had better prediction accuracy for SPAD value of leaves, R² and RPD were 0.93 and 2.19, respectively. NIR spectroscopy with chemometrics was feasible to identify the citrus diseases on physicochemical indexes of leaves. The finding can provide an important reference for the real-time field detection of citrus diseases.