For information in English, please click here
Research on cotton water content prediction based on computer-vision
Abstract：This paper aimed at establishing a rapidly, non-destructively and low-costly model to predict cotton water content in different growth periods based on computer-vision technology，for providing theoretical basis for irrigation decision-making. Cotton canopy digital image under different drought stress were captured by Nikon D90 digital camera during the period of cotton initial bloom stage, full bloom stage and flowering and boll-setting stage, respectively. The color characteristic parameters of the images of cotton canopy were extracted by a image processing software. Green parts of cotton image were extracted at hue between 60-180 degree. Then RGB and HSBr value of cotton canopy image were extracted. Meanwhile, the cotton water content for each sample was detected using the oven dry method in the laboratory. The result showed that the correlation coefficient between color characteristic parameters and cotton water content reached extremely significant level in different growth periods. The cotton water content predictive models at different growth stages were established separately by using the stepwise regression method. The color characteristic parameters of G/R in cotton initial bloom stage; S/(B/R)and S/(2G-R-B) in full bloom stage and 2G-R-B in flowering and boll-setting stage with cotton water content established the best prediction model respectively. The correlation coefficient of the predict model were 0.886, 0.890 and 0.831 in three different periods. The collected samples were used to construct a large-scale data set，two-thirds of which were used as the train set and the remaining one third were used as the test set. In order to test the prediction accuracy of models, the determination coefficient (R2) and root mean square error (RMSE) between plant water content of the measured values and predicted value were calculated by mathematical methods. The water content predicts model in initial bloom stage had highest accuracies, the R2 reached to 0.874 and root mean square error was 1.476. The model in full bloom stage had higher accuracies, the R2 reached to 0.602 and root mean square error was 1.096. The model based in flowering and boll-setting stage had lower accuracies, the validation R2 reached to 0.790 and root mean square error was 0.956. The relative error between predicted values and measured values were less than 2% in different periods, and the predicted results were reliable. The data of different growth periods was merged, established the universal model suitable for predicting plant water content at different stages .After accuracy test, universal model prediction accuracy was higher than the best model of initial bloom stage , and less than the best model of full bloom stage. The root mean square error between predicted values and measured values was 1.096 and relative error was 1.40%. It was concluded that color characteristic parameters of the canopy images can predict cotton crop water content based on computer-vision under the conditions of field natural light.
Click to fold