| How to detect plant disease rapidly and accurately is an important research filed in precision agriculture. Althouth biochemical methods can identify specises of pathogens, it can’t meet the demand of real-time detection in agriculture. This study utilized QE65000spectrometer to examine the potentials of non-imaging sensor systems for the detection of gray mould of peppers and tomatoes, which represents solanaceae vegetables. Combined with common vegetation indices, data mining technique and phytopathology knowledge, different models were established and analyzed. In addition, a new two-waveband portable canopy analyzer was designed for further study and appilication. The main results were achieved as follows.(1)Spectral reflectance experiment of plant canopy must be made in uniform illumination environment. For hyperspectral imaging platform that is difficult to be moved to outdoors, it is possible to stimulate uniform illumination in small scale with many (maybe over8) halogen lamps. It is not necessary to put white panel at the same height with the top of the plant canopy when illumination calibration is operated, because the size of white panel is often much smaller than the canopy. The experiment result showed that the change of distance between the white pannle with the detector probe did not influence the calibaration result in the scale of50centimeters. In the regions and seasons with sunny weather, ASD series spectrometer is more appropriate to be used in the outdoor experiment, while in the regions and seasons with rainy weather, QE65000spectrometer is better.(2)A kind of pepper (Japan Fuji pepper) and two kinds of tomato (Fen Oubao&Zheza809) at late seedling stage were used as the host of gray mold and detected at the canapy scale. In order to make it easier to apply the results to equipment design and compare the results of diffirent researches, the study utilize only raw spectral data and preprocessing data with smoothing agrithom. Full spectrum partial least squaress discriminant analysis (PLS) models were built for3kinds of host-gray mold systems. The whole forecast accuracy rate was78.5%(pepper),84.3%(Fen Oubao) and82.4%(Zheza809) with the threshold value of±0.5. After the propocessing, only the accuracy rate of Zheza809group was imporved to87.1%. Based on PLS model, this study utilized loding weight method and regression coefficient method to select characteristic wavelengths. It was found that the wavelengths of three different groups with two methods showed good consistency.(3)After plants were infected with diseases, the inner pigment, water content, nutrient and so on would be changed. Spectral vegetation indices were widely applied in the remote sensing and plants pigement and nutrient detections. Therefore34common indices were used and selected according to the Pearson coefficient that is larger than0.5. Then with different input amounts of14(pepper),20(Fen Oubao),17(Zheza809),3models of least squares support vector machine (LS-SVM) were estabilised. The whole forecast accuracy rate was79.4%(pepper),83.1%(Fen Oubao) and87.1%(Zheza809) with the threshold value of+0.5, which was higher than the outcome of PLS models. In order to validate the model effect of vegetation indice, full spectrum LS-SVM models were established for individual group, with the prediction accuracy of82.8%(pepper),88%(Fen Oubao) and88.2%(Zheza809). The input dimensions of LS-SVM model decreased by97.5%, but the precision of model decreased a little. As a result, using vegetation indices is a valid way to extract spectral information.(4) For the purpose of improving the stability and universality of models, integrated the results of3experiments,13kinds of indices that were strongly correlated with canopy gray mold were chosed as follows:GNDVI, OSAVI, OSVAVI2, MTCI, Mac, ND705, PSNDb, Datt, mSR, PSNDb, MTCI, SIPI and PSSRb.3separate LS-SVM models were built with13indices as model input and the whole prediction accuracy was74.2%(pepper),89.2%(Fen Oubao) and85.5%(Zheza809). By pair wise comparison of13indices, the model input was decreased to9indices and7indices, which the redundancy of model input was further reduced. And the model result was declined from84.9%(pepper),75.9%(Fen Oubao) and84.3%(Zheza809) to58.1%(pepper),73.5%(Fen Oubao) and74.7%(Zheza809). By analyzing the model results of different input amounts, it was concluded that model prediction accuracy was satisfactory with9to13indcies (listed as above). Finally, based on all samples (Zheza809), two band normalized difference indice (Rλ1-Rλ2)/(Rλ1-Rλ2) and two band simple ratio indice Rλ1-Rλ2was created (λ1=732.58nm, λ2=549.22nm). Their correlation coefficient with gray mold at canopy scale was0.8257and0.8227, separately. Meanwhile, both indices was validated with another two groups of experiments. Althouth the correlation coefficient was smaller than0.8, the results was close to or beyond the highest coefficient of common indices. The results proved the stability of both indices and provided reference for further studies.(5) A canapy analyzer with two bands that detects gray mold of solanaceae vegetables was studied and designed. As a result of gun-type plastic casing, the design of light path adopted bifurcation fiber. Self-designed light channel fixed four light filters and achieved fast replacement of filters. The portable equipment utilized wireless way to transfer data. Control program on laptop was written based on Visual Basic language, achieving data vision, curve plotting and data storage in Microsoft EXCEL. Also the program includes a linear model based on GNDVI to indentify healthy samples and samples infected with gray mold. After preliminary test, the equipment could provide comparatively precise results. |
PDF Full Text Request