基于近红外光谱技术的果树花期树种识别方法

Automatic detection of fruit tree species during the flowering period using near infrared spectra technologies

为建立果树花期树种识别的有效模型,利用ASD Field Spec 3全波段便携式光谱分析仪采集了4种果树花期花的光谱数据。利用剔除异常光谱、5点移动平滑等技术对4种果树花期花的光谱反射率进行预处理,使用连续投影算法(SPA)进行有效波长选取并获得7个波长下的反射光谱,同时增加了590 nm和720 nm处2个波形差异大的光谱,与归一化植被指数(INDV)和比值植被指数(IRV)共11个特征波段作为分类建模数据,建立了偏最小二乘判别分析(PLS-DA),正交偏最小二乘判别分析(O-PL-DA)和基于误差反向传播算法的多层前向神经网络(BP)算法3种识别模型。结果表明:对测试样本的识别率由高到低依次为BP(93.90%)>O-PLS-DA(81.82%)>PLS-DA(76.36%)。综合研究认为:在优选波段的基础上,对果树花期树种判别应优选BP神经网络模型。

To establish an effective model for fruit tree species identification at the flowering stage, spectral data of four kinds of fruit trees were collected using an ASD Field Spec 3 full band portable spectrometer. Nine sensitive and characteristic bands of the spectrum（370 nm, 395 nm, 541 nm, 590 nm, 682 nm, 720 nm, 1 268 nm, 1 839 nm, and 2 481 nm） and two vegetation indices for accurately detecting fruit tree species were first obtained using the Successive-Projections-Algorithm（SPA） method. Subsequently, some classification methods were applied, such as Partial Least Squares Discriminant Analysis（PLS-DA）, Orthogonal Projection to Latent Structure Discriminant Analysis（O-PLS-DA）, and Back Propagation（BP）, to compare their effectiveness for distinguishing fruit tree species. Choice 30 m × 30 m standard rural area in the garden, select 10 trees of every4 fruit tree species and every tree select 3-4 points using optional bolting method. 10 spectra were measured and take the average at every point, repeated three times. Results showed that the average detecting accuracy for PLS-DA was 73.36%, for O-PLS-DA was 81.82%, and for BP was 93.90% with the BP model having the best prediction accuracy for clarifying fruit tree species. This study demonstrated the feasibility of implementing hyperspectral imaging from near infrared spectra technologies（NIST） for identifying fruit tree species during the flowering period.