Study On Cotton Drought Diagnosis Based On Computer Vision

This study was proposed to establish a computer-vision based on drought diagnosis model by cottoncanopy digital image analysis. Cotton canopy digital images were taken by digital camera under differentdrought stress. The color gray-scale value of cotton canopy image was extracted by the computer visiontechnology. At the same time, the change of cotton physiological indexes, cotton plant water content andsoil water content under drought stress were monitored. Correlation between cotton physiologicalindexes, cotton plant water content, soil water content, leaf chlorophyll content and color gray-scalevalue obtained from cotton canopy digital image was studied, and the best fitted models were set-upthrough stepwise regression eventually. The study also studied the vertical projection area of cottoncanopy, which get from digital image of potted cotton, as a diagnosis tools for cotton drought diagnosis.In all those studies, gray board was adopted to adjust image at different growth periods and made cottoncanopy images captured in different growth period can be normalized and in a single model. The mainconclusions are as follows:(1) Drought stress resulted in a series of change of cotton leaf proline content, soluble protein content,MDA content and soluble sugar content, and whole-plant water content. The results showed that pottedor field cotton plants leaf proline content, soluble protein content, MDA content and soluble sugarcontent were increased with the increase of the degree of drought stress, with prolonged drought stressincreased. Potted cotton and field cotton plant water content decreased with the increase of the degree ofdrought stress. Cotton leaf physiological indexes of proline, soluble protein, MDA and soluble sugarwere significant negative correlation with water content. Cotton plant water content can monitor cottondrought stress as an indicator.The color characteristic value of cotton canopy image was extract by software and researched thecorrelation of potted cotton and physiological indexes. Then the prediction models wereestablished based on different stages physiological indexes. The best predicting models of potted cottonleaf proline and MDA content can be established with the color grayscale saturation(S) value under budstage drought stress. The best predicting models of potted cotton leaf soluble protein content canbe established with the color characteristic r-b(standardized red and blue) value and soluble sugarcontent predicting model can be established by g-b(standardized green and blue) value under bud stagedrought stress. The best predicting models of potted cotton leaf proline, soluble protein, soluble sugarand MDA content can all be established with the color grayscale blue(B) value under beginning-flowerstage drought stress, and its determination coefficient of equation was greater than 0.8.The best predicting models of potted cotton plant water content in bud and beginning-flower stage canbe established with the color characteristic S value and its determination coefficient of equations were0.789 and 0.623, respectively.(2) Gray board was proposed in this study to adjust digital image brightness vibration in differentgrowth period due to temperature and solar altitude differences. The average value of bright(Br) was setas the criterion value to correct the brightness value of different growth stage, and then the brightnessgray-scale values of images in the different period were normalized. The HSBr values after gray boardcorrection value is converted to RGB values, then gray board correction algorithm of color grayscalevalue from cotton canopy image were obtained eventually. The forecasting models of plant water contentbefore and after grey board correction were established respectively.The best predicting models of field cotton water content can be established with the colorcharacteristic G/R which before gray board correction at beginning-flower stage drought stress, its Rvalue was 0.900. The best predicting models of field cotton water content canbe established with the color value G/(R+B) after gray board correction under full bloom stage droughtstress; its R value was 0.887. And under flower and boll-setting stage drought stress, thebest predicting models of field cotton water content can be established with the color characteristic B/Gwhich was after gray board correction, its R value was-0.861. After examination, the measured values ofoptimal prediction model at different periods of plant water content between the predicted and the rootmean square error is 1.476, 1.044 and 1.065 respectively, the relative error less than 2%, the model hasgood stability, can well predict plant water content in different periods. Validation test demonstrated thatbetween the predicted values calculated by the optimal prediction equation at different periods and themeasured values of cotton water content, the root mean square errors(RMSE) were 1.476 、 1.044and1.065 respectively, and the relative errors(RE %) were all less than 2%. The predicted models canpredict cotton water content in different periods with a sound accuracy and stability. A general model topredicted cotton water content suitable for different growth stages was established by combined data indifferent periods. The prediction accuracy of the general model after grey board correction was betterthan the model which was established by original data. The best general prediction model wasy=52.023+0.7666(G-R), which were established by data corrected by grey board. The decisioncoefficient of the model was 0.782. Validation test showed its root mean square error(RMSE) was 2.03,and the relative error(RE %) was 1.79%. Gray broad adjusted model was better than model derived fromorginal data.(3) The optimal predicting models of potted cotton soil relatively water content canbe established with the image’s grayscale b and S value at bud and beginning-flower stage, respectively.The determination coefficient of predicting models were 0.637 and 0.684(p<0.01), respectively.The best predicting models of field cotton soil relatively water content can be established with the colorcharacteristic G-R、B/G and G-R value before gray board correction under beginning-flower, full bloomand flower and boll-setting stage, respectively. The best predicting models of field cotton soil relativelywater content can be established with the color characteristic G-R、G/(R+G+B) and G-R value after grayboard correction under beginning-flower, full bloom and flower and boll-setting stage drought stress,respectively. All the models achieved extremely significant level. After inspection, the results displayedthat model prediction accuracy after grey board corrections were higher than before grey boardcorrection. The best general predicting models of field cotton soil relatively water content canbe established with the color characteristic(G-R)/B after gray board correction. Between the predictedvalues which were calculated by the best general prediction model and the measured values of soilrelatively water content, its root mean square error(RMSE) was 10.75, and the relative error(RE %) was17.62%. Grey board correction can improve the precision of soil relative water content prediction modelin the each growth and general period.(4) Potted cotton plant absolute canopy projected area can direct the size of cotton canopy leaf areaand biomass. Relative projected area of cotton canopy reflected the actual intensity degree changes ofcotton canopy. In the process of continuous monitoring potted cotton plants during the flower andboll-setting period, the relative canopy projection area of potted plant cotton leaf and photosynthetic rate,transpiration rate and stomata conductance have the same change trend. Relative canopy projection areaand different treatment of soil relative soil water content achieves significant correlation. Aftercalculation, the relative projection area reduced to 0.63-0.67 can be used the threshold of the potted plantcotton soil relative water content dropped to 50%.(5) A significant correlation can be observed between chlorophyll content and color indexes, and thesetrend keep in accordant in different growth period. Using the gray board average brightness of differentrecording date as the normalized calibration standard, merging different date color characteristic valueand chlorophyll content, the chlorophyll prediction model were established by original and gray boardadjusted, respectively. The correlation coefficients between original DGCI and R-B and cottonchlorophyll were 0.8857 or-0.8726, and 0.9073 or-0.9016 in corrected data, respectively. Thecorrelation coefficient between parameters and chlorophyll content were improved after grey boardadjustment. The color characteristic parameters and chlorophyll content in different periods werecombined, and the correlation between them was analyzed. DGCI and R-B had the most significantlinear correlation with cotton leaf chlorophyll content. Comparing the chlorophyll content predictionaccuracy of DGCI or Red-Blue by original or adjusted data, it showed that the parameter DGCI orRed-Blue after adjustment model prediction accuracy is higher than original data. The predictionaccuracy of DGCI is higher than R-B after calibration. Between the predicted values, which werecalculated by the equation, and the measured values of chlorophyll, its root mean square errors(RMSE)was 0.1200, and the relative errors(RE %) was 5.28%. The decision coefficient was 0.8812. Theprediction accuracy was better. Our results demonstrated that the adjusted DGCI was the best indicator topredict cotton leaf chlorophyll content, and the prediction model was feasible for applying computervision technology to rapidly predict cotton leaf chlorophyll content.