A Study On Growth Monitoring Of Maize In Field Using Image Processing

Maize,as one of the most cultivated cereal crop all over the world,is widely used in food,feed,and industrial raw materials.Although mechanized agriculture have greatly improved agricultural productivity,there are some environmental issues which threat the interests of farmers.As a result,development of precision agriculture has become the core of digital agriculture.Automatic monitoring crop growth states plays a vital role in the development of crop production and automatic management.Compared with artificial observation,machine vision and image analysis based on image processing technology have the merit of informative,process fast and high precision.They have great potential in saving labor force and reducing subjective influence of artificial observation.At present,image-based machine vision and image analysis technology have been widely used in agriculture,targets are usually recognised by analyzing the shape,texture,color and spectral information of weeds,crops and soil obtained in field images.Judgment and prediction of growth in real time can help people analyze the relationship between crop growth status and agricultural meteorology,and provide effective agricultural assistance to increase agricultural production.In this paper,a real-time,lossless,accurate and robust automatic monitoring algorithm for maize growth is developed based on machine vision and pattern recognition.The research priority is automatic identification of maize,weed and tassels using computer vision and using coverage information of the image with farmland microclimate information to establish the maize growth identification and prediction model.The main research contents include:1.Maize and weed classification based on support vector data description(SVDD).A total of 396 maize field images were collected during 44 days of three years from seedling stage to jointing stage.Using threshold segmentation algorithm with ExG feature of each pixel to extract the green vegetation area,then extracting the vegetation index of each area to establish SVDD model after feature dimension reduction by PCA.The experiment results show that using the vegetation index and SVDD algorithm can effectively identify the field image of maize and weed area,and average accuracy of three years were 90.19%,92.36% and 93.87%.The results are robust to different years and different areas.It provides a possible technical approach for automatic identification of maize and weed in the field.2.Automatic identification of maize tassel based on saliency detection.This work proposes a novel method for automatic tassel detection.In the algorithm,a light saturation removal model was used by modeling the scene depth of saturation graph to remove image saturation.Then Itti visual attention detection algorithm was used to detect the area of interest.Finally,texture features were used to develop a classification model to eliminate false positives.Experimental results show that the proposed method can effectively reduce the effect of light saturation and improve the recognition accuracy greatly,three indicators increased by 3.06%,10.17% and 6.44%.And it is better than SIFT and SLIC algorithm with 3.04% and 1.45% in F1 features respectively.The results indicate that this proposed method can effectively detect maize tassels in field images and remain stable over time.3.Maize growth prediction based on multi-information fusion.In this paper,the coverage of maize in three years was used to combine the information of farmland microclimate to predict maize growth period using T-S fuzzy modeling.The T-S fuzzy identification model was established by using historical coverage and farmland microclimate information in 5 days.Then the 5th day’s prediction coverage information was combined with farmland microclimate to predict the growth stage of maize.The experimental results show that the T-S fuzzy model can effectively simulate maize growth coverage.And prediction accuracy is improved by combining farmland microclimate information with coverage information,and prediction accuracy of three years were 98.78%,94.51% and 96.20%.It provides a possible technical approach for the growth monitoring of maize in field.