Abstract: Abstract: Timely, accurate, rapid diagnosis and grading of citrus Huanglongbing (HLB), a devastating disease severely influencing the citrus industry in the world, plays a very important role in guaranteeing the yield, the quality of citrus fruits, and the benefits of citrus growers. Based on a hyperspectral imaging technique, this paper not only focused on the method of early nondestructive detection and grading of citrus HLB disease, but also tried to discuss the influence of different data preprocessing methods on the modeling results. What is more, the varying reflection spectral characteristics of citrus leaves in diverse disease degrees were analyzed in the paper based on measured hyperspectral data.Hyperspectral images of five kinds of citrus leaves, including the healthy, infected with different degrees with HLB, and those with zinc deficiency, were acquired through experiments by a hyperspectral imaging system with the wavelength range of 370-1 000 nm, and then the average spectral reflectance data of region of interests of different kinds of leaf samples were obtained by utilizing the environment for visualizing images(ENVI). By taking advantage of a partial least squares-discriminate analysis (PLS-DA) method, three models of the early diagnosis and grading of HLB disease, tested with a leave-one-out cross-validation strategy, were established with original spectral data and data preprocessed by different data pretreatment methods, such as first derivative and moving window polynomial fitting smoothing (Savitzky-Golay smoothing,SG). In the end, the predictive performances of all of the three models were compared and analyzed with the new validation data.As a result, the cross-validation correlation coefficients of three discriminate models were all greater than 0.9548, however, their prediction performances were not the same. The detection results of the first discriminate model, established with original data, was not satisfactory. The second discriminate model, set up with data pretreated by a first derivative method, could basically identify the three types of HLB correctly: the mottled, yellowing, and no obvious symptoms, but there were quite a few healthy samples and zinc deficiency samples misjudged as HLB disease. What was satisfying was that the third model established with spectral reflectance data preprocessed by the Savitzky-Golay smoothing and first derivative methods had the best discriminate effect, which achieved prediction accuracy of no less than 92% of five kinds of leaf samples, the overall classification accuracy rate was 96.4% (in a test set of 250 samples, 241 having been correctly indentified), as well as RMSEP of 0.1344. In addition to these, its prediction accuracy for the healthy, zinc deficiency samples were 92% and 96%, which meant that there were still a few samples having been mistaken for HLB disease. As for the unobvious symptom, the mottled, and the yellowing samples, although some wrong judgments still existed among them, at least all of the three types could be correctly identified as infected with HLB. No matter what, the above research results showed that this method for early, nondestructive diagnosis of citrus HLB was of great significance and feasibility. The research was able to provide a new method for early detection and pre-warning of citrus disease, and also lay a basis for remote sensing monitoring of HLB disease degrees.