Fine Extraction Of Plateau Mountain Tobacco Plants Based On U-Net And Superpixe

Tobacco is the main cash crop in China.China’s tobacco industry plays a vital role in national and local economic construction.Especially in the plateau and mountainous areas of southwestern China,planting tobacco leaves is an important means to help the poor get rid of poverty and increase farmers’ income.Tobacco refined management is particularly important in tobacco planting.Accurate control of the number of tobacco plants and spatial distribution information can provide a basis for later fertilization,irrigation,and pest control,as well as accurate estimates of tobacco leaf yield.At present,the statistics of tobacco plant information are still carried out by manual field surveys,but the tobacco planting areas are mostly plateaus and mountainous areas,with large terrain fluctuations,and the planting areas are relatively large and relatively scattered.This method has low work efficiency and strong human subjectivity.In order to reduce labor costs and improve the efficiency of tobacco plant counting,this paper proposes an automatic extraction and counting method for tobacco planting surfaces and tobacco plants based on UAV visible light band images,combined with deep learning algorithms and image processing technology.Main research contents:(1)Extract tobacco planting surface based on U-Net and connectivity enhancement algorithm.Aiming at the shortcomings of traditional object-oriented methods,such as low automation degree,low extraction accuracy and cumbersome extraction process,an automatic extraction method of tobacco planting surface based on deep learning was proposed.Using UAV image data to make a dataset,and performing normalization,data enhancement and other preprocessing,using U-Net semantic segmentation network to train and predict the dataset,and propose a method to remove small patches and fill holes for connectivity enhancement The algorithm corrects the prediction result and realizes the automatic extraction of tobacco planting surface.The results show that the proposed method can effectively extract tobacco planting surface,and its segmentation index mean Intersection over Union(mIoU)is98.24%;(2)Tobacco planting rows were extracted by fusion vegetation index.In order to finely extract tobacco plants and eliminate the interference of background objects on the extraction of tobacco plants,an Excess Green Difference Index(EGDI)was constructed according to tobacco spectral information.Based on the results of tobacco planting surface extraction,combined with the characteristics and advantages of Excess Green(EXG)and excess green index,other non-tobacco ground features were removed,and tobacco planting rows were extracted to facilitate subsequent tobacco plant extraction;(3)Extract tobacco plants by combining superpixel and distance transformation algorithm.On the basis of tobacco planting row extraction,the simple linear iterative clustering algorithm(SLIC)is used to perform superpixel segmentation on tobacco planting surface.Aiming at the misclassification phenomenon in the segmentation,a distance based on Euclidean distance and distance based on coordinate components is proposed.The transformed algorithm corrects the incorrectly classified tobacco plants,and finally realizes the extraction and counting of tobacco plants.The results show that the number of tobacco plants finally extracted by the two distance transformation methods has little difference with the actual number,the average overall accuracy is above 97%,the misclassification and omission are relatively few,and the average accuracy is greater than 94%.The method has a high degree of automation,and the extraction accuracy can meet the actual production management requirements.