Research On Estimation Of Lettuce Growth Index Based On Machine Vision

Factory crop cultivation is the highest stage of facility horticulture development,which can effectively solve the problems faced by traditional crop cultivation such as insufficient arable land,insufficient yield and great environmental impact.Because of its short growth cycle,high nutrient content,and high yield,lettuce has become one of the leafy vegetables with the largest industrial production volume at this stage.Using digital images to accurately,quickly and non-destructively monitor the growth indicators of lettuce in different growth periods in factories is of great significance for monitoring the growth of lettuce,judging the accumulation of nutrients,improving growth efficiency and quality,and determining the best harvest period.In order to better realize the dynamic monitoring of the lettuce growth process,this paper takes the hydroponic lettuce with different growth characteristics commonly used in factories as the research object,and obtains digital image data and corresponding indicators of different varieties of lettuce in different growth periods and different scales.Data,analyze the correlation between multiple growth indicators(leaf area,fresh quality and chlorophyll content)and the color,shape and texture features of the images,use data mining algorithms to screen the optimal feature subset,and combine machine learning models to build different varieties Quantitative estimation models for different growth indicators at different scales,evaluating the application and prediction accuracy of digital image data in lettuce growth indicators.The main research contents and conclusions are as follows:(1)Object extraction for lettuce images of different scales in the factory environment.Firstly,the image is clipped,and for the canopy image,the distribution of pixels is analyzed by using the 9-channel histogram of the RGB,HSV and L*a*b* three color spaces of the image,and the most suitable target is found by combining the image characteristics of different channels.The segmented a* channel is subjected to iterative threshold segmentation and morphological processing to obtain a binary image with complete segmentation of the target and background,which is converted into a binary image and then matrix multiplied with the original color image to obtain a complete lettuce at different scales.The color segmentation image overcomes the interference of light,shadow and irrelevant impurities;for the side view image,using a special normalized rgb color space,this method can effectively highlight the green target,and combine the corresponding green pixel value to extract the target.(2)Extraction of features related to different growth indicators.For the leaf area and fresh quality indicators,594 single-plant scale and 320 population-scale image data were analyzed using Python-Open CV software,and 21 color features,8 shape features,and 18 texture features were fully extracted,and a set of image evaluation indicators consists of 47 global feature variables.For the chlorophyll content index,144 leaf-scale and 144 single-plant canopy-scale image data were analyzed respectively,and 29 color features were extracted comprehensively,and a set of image evaluation index sets integrating multiple color spaces was established.(3)Image feature index set screening and modeling analysis.Aiming at the characteristic index sets at different scales obtained by different growth indexes,a group of optimal features related to the indexes are screened by using the Pearson coefficient method + forward stepwise regression model and gradient enhanced decision tree(GBDT)+ forward stepwise regression model Subsets,further use methods such as linear regression(LR),support vector regression(SVR),random forest(RF),ridge regression(RR)and gradient boosting regression(GBR)to construct a quantitative estimation model of lettuce growth indicators,through nested crossover.The verification and parameter optimization determine the optimal feature screening method and the optimal model for different indicators,and then use the verification set data to verify the generalization ability of the optimal model.The results show that the optimal feature selection algorithm and the optimal model of leaf area and fresh quality index are both the combination of GBDT+GBR.Among them,the R~2 of the leaf area index on the single-plant sample test set of ‘Lvluo’ was 0.9715,and the R~2 of the ‘Bixiao’ single-plant sample test set was 0.9072.The R~2 of the fresh quality index on the ‘Lvluo’ sample test set is0.9464,and the R~2 of the ‘Bixiao’ sample test set is 0.8994.For the chlorophyll content index,the optimal feature selection algorithm and optimal model at the leaf scale are the combination of GBDT+GBR,and the optimal feature screening algorithm and optimal model at the canopy scale are the combination of GBDT+RF.Among them,the test set R~2 at the leaf scale is 0.8267,the RMSE is 2.5431,the verification set R~2 is 0.8206,and the RMSE is 2.6227;the test set R~2 at the canopy scale is 0.8128,the RMSE is 2.9667,the verification set R~2 is 0.8027,and the RMSE is 2.9831.