A hyperspectral GA-PLSR model for prediction of pine wilt disease

Pine wilt disease caused by a forest-invasive alien species, the pine wood nematode (Bursaphelenchus xylophilus) is considered as one of the most destructive pest problems. In recent years, spectroscopic technologies have shown great potentials for the assessment of forest damage due to their nondestructive, noninvasive, cost-effective, and rapidly responsive nature. This paper first identified the hyperspectral characteristics of pine wilt disease by measuring and analyzing the changes in spectral reflectance of healthy and infected Pinus massoniana trees. Then 16 spectral features were extracted from the spectral bands covering the green region (510~580 nm), the red region (620~680 nm), the red edge (680~760 nm), the near-infrared region (780~1100 nm), and coded as genes composing the chromosome of a genetic algorithm (GA). Based on the optimal spectral features with suitable fitness from the GA, a partial least squares regression (PLSR) prediction model was built with highest determination coefficient R2c = 0.91, R2v = 0.82, relative prediction deviation RPD = 3.3 and lowest root mean square error RMSEc = 0.23, RMSEv = 0.33 on the calibration and validation datasets. Compared with other PLSR models, our proposed GA-based approach significantly improves the prediction accuracy with few input spectral features.