Diagnosis Of Water Deficiency In Watermelon Seedlings Based On Multi-source Information Fusion

In seedling production,water and fertilizer management are mainly carried out manually,resulting in serious waste of manpower and resources.The water condition of the seedlings will cause the change of the color,texture,shape,temperature and other characteristics of the leaves.In this study,three varieties of watermelon seedlings of "Zaojia84-28","Early spring carbanyu" and "Mutian Hongxing" were used as experimental materials,and computer vision technology was used to detect these characteristics to establish a water content prediction model and integrate it into an expert system,which could realize rapid and accurate water monitoring and irrigation decision-making.The specific research contents and conclusions are as follows:(1)Considering that the traditional visible light segmentation method cannot be applied in complex environment and is susceptible to noise,a semantic segmentation model based on UNet is proposed to segment the whole plate of watermelon seedling color image,and VGG16 is used as feature extraction network to optimize the traditional UNet segmentation model.The average pixel segmentation accuracy of the optimized VGG16-UNET model is99.22%,0.76% higher than that of the traditional UNet model.The average crossover ratio reached 96.56%,1.91% higher than that of the traditional UNet model.This algorithm can perform image segmentation of watermelon seedling canopy well.(2)The edges of the infrared thermal image of the whole disk of watermelon seedlings were uneven and the gray level of the image changed sharply.The traditional threshold segmentation method could not accurately segment the near-infrared thermal image of the canopy.Ilf camera calibration method is used to calculate the thermal and visible light cameras inside and outside the parameters of the camera,and then based on the internal and external parameters to calculate the thermal image to visible light image projection transformation matrix,completed by projection transformation matrix thermal image and visible light image alignment,with the final alignment color figure canopy in the reference image segmentation,The target canopy region of infrared thermal image of watermelon seedling was extracted.The extraction effect of canopy area is good.(3)Aiming at the problem of insufficient accuracy and robustness of single information source modeling,a water content prediction modeling method based on multi-source information fusion was proposed.A total of 68 features of temperature,color,texture and phenotype of watermelon seedlings were extracted after segmentation,and Pearson correlation coefficient and P-value significant coefficient between each feature and water content were calculated respectively.Then 29 features with high correlation and significant performance were screened out.This feature extraction method effectively ensures the feasibility and accuracy of water content prediction model.(4)Use RandomForest regression model,CatBoost regression model,XGBoost regression model,Voting integration regression model and Stacking integration regression model respectively to establish regression prediction model of watermelon seedling water content.The determination coefficient R2 and root mean square error RMSE of each model are calculated.Finally,by comparing the prediction results,it was determined that the Stacking integration regression model had the most accurate prediction effect and the best fitting degree.R2 and RMSE were 0.9125 and 0.005820 in one-leaf single-heart period,R2 and RMSE were 0.9223 and 0.005126 in two-leaf single-heart period.R2 and RMSE were0.9065 and 0.006425 respectively.The overall accuracy is very high.(5)Combined with planting expert knowledge and water stress experiment,the average water content of watermelon seedlings in each growth stage under critical water shortage condition was determined,and then combined with watermelon seedling water content prediction model,an intelligent water shortage diagnosis system was established using Py Qt5,and the feasibility and practicability of the system was verified through experiments.Extraction and this study combines the infrared thermal imager,color cameras and depth camera to obtain images of multi-source information,set up the precise model for predicting water cut watermelon seedlings,and in combination with the determination of the test each growth stage lack of critical moisture content of watermelon seedlings lack of diagnosis system is established,provide decision-making basis for greenhouse watermelon seedlings online irrigation.This study effectively promoted the development of nondestructive testing technology of crop water and laid a technical foundation for fine management of water in different growth stages of watermelon seedlings.