Study On Application Of Spectral Imaging Technique In The Citrus Harvest And Post-harvest

Harvesting and post-harvesting are two necessary stages for citrus from farm to market, which great attention shall be paid on, so that the highest benefits can be achieved with the minimum capital invested. As the gradual reduction of agricultural labor force, long time-consuming and labor-intensive operation of manual picking, and also the easy-caused damages in fruit by traditional ways of picking (e.g. hand picking, scissor cutting, pole stabbing), which affects fruit quality largely, citrus harvesting robot is in urgent need to reduce the labor input, as well as ensure fruit quality and improve production efficiency. Post-harvesting treatments of citrus mainly include cleaning, waxing and grading. At present, the technologies of cleaning and waxing have been widely applicated in fruit on-line processing. However, the fruit still not been graded based on their quality. Fruit, not graded or not fully graded, leads to grading confusion, which can not realize the value of superior fruit and the efficient increment of the revenue.Correct recognization of the object is the primary issue to be addressed for picking robot, since errors at this stage may affect the future processing all the way. However, the non-structural natural environment (e.g. uncertain light conditions, complex backgrounds, fruits diversity and random growth of fruits) will cause serious difficulties for traditional recognization methods, which based on intensity, color or shape. Thus, spectral imaging technique was proposed to recognize fruit in a tree canopy under the unstructured environment.The grading technology is rather behind in China, especially, there are few literatures about detection of citrus rust. Citrus rust is one of the most important indicators of the gradation, and the color is similar between the surfaces of rust and normal citrus. Therefore, spectral imaging technique was suggested to detect citrus rust, meanwhile, data mining methods for spectral imagery were explored.Major progresses in this study are as follows: (1) Recognition of citrus in the nature scene using spectral imaging techniqueFirst, different bands images in the same scene were acquired by a self-designed filter-based spectral imaging system to build spectral image data cube; then, the spectral image datacube was preprocessed by MNF(Minimum Noise Fraction); finally, SAM(Spectral Angle Mapper), MF(Matched Filtering) and LSU(Liner Spectral Unmixing) were implemented comparatively to extract the citrus features. Experiment results show that SAM has the best overall classification performance, with success rate of 96.3% and error rate of 2.4%, followed by MF(94.3%,4.3%) and LSU (93.3%, 4.9%). This work demonstrates that spectral imaging technique can be used effectively to recognize citrus in the nature scene. Furthermore, spectral imaging technique is superior to the traditional recognition methods. SAM is able to effectively reduce the adverse effects of uneven reflectance intensity, and strongly resist the molest of complex background. The recognition accuracy of SAM is the highest, as well as the robust. (2) Detection of citrus rust using spectral imaging techniqueFirst, the spectral imageries of citrus samples (including normal and rust samples) were acquired by a self-designed spectrometer-based spectral imaging system; then, imagery correlation and some other preprocessing methods (e.g. calibration) were achieved to scale the noise and improve the subsequent spectral processing results; finally,12 methods (e.g. the supervision and non-supervised classification based on feature space, spectral matching technique based on spectra space, twice-steps principal component analysis and band ratio method) were adopted respectively to detect citrus rust. Experiment results show that the detection accuracies of SAM and MDC(Mahalanobis Distance Classification) were the highest (96% and 94%, respectively), followed by the band ratio method (92%), BE(Binary Encoding) (92%), two-step principal component analysis (90%) and MLC(Maximum Likelihood Classification) (90%). The detection accuracies of the remaining methods were all lower than 90%. This work demonstrates that the non-supervised classification (K-Mean and ISODATA) based on feature space, are infeasible for detecting citrus rust. In contrast, MDC, SAM are less impacted by the selection of the training set, its algorithm is simpler and more precise. But SAM and MDC process with whole preprocessed data, which will take more computing time, are not suitable for the online detection. Due to twice-steps principal component analysis and band ratio method have optimized the characteristic wavelengths, the amount of calculation are significantly reduced. The characteristic wavelengths of 571,652 and 741nm obtained by twice-steps principal component analysis or 621,625 and 717nm obtained by band ratio method could be used to design a filter-based spectral image system in order to achieve rapid and non-destructive online detection of citrus rust.The application of spectral imaging technique in the citrus harvest and post-harvest was studied in this paper. It reached to a national and even international advanced level in related fields. Research topics has a certain reference value for the recognition of picking harvest and fruit gradation, and also has far-reaching economic significance for improving fruit quality, economic efficiency and the international competitiveness of national fruit industry, meanwhile, lay the foundation for the application of spectral imaging technique in related fields of agricultural products.