Deep learning based automatic acquisition of plot-scale seedling deficiency information in sugarcane fields

【Objective】 To improve the degree of automatic acquisition and accuracy of sugarcane fields seedling deficiency information, which could provide reference for the realization of automated extraction of sugarcane field seedling deficiency location information in plot units. 【Method】 Images of sugarcane fields in the study area were collected by unmanned aerial vehicle（UAV） and a dataset was created to detect and recognize sugarcane seedlings using the YOLOv8model. The recognition results were vectorised,rotated and clustered to accurately calculate the number of rows and spacing information. Finally,the vector point distribution maps of sugarcane seedlings in the ideal situation of no seedling deficiency in the field and the actual distribution maps of seedling deficiency locations were produced to assess the overall seedling deficiency situation in sugarcane field.【Result】The YOLOv8 model identified sugarcane seedlings with an accuracy of 98.84%, a recall of 90.76%, and an average precision of 97.05%. Even in environments with high weed distribution, where weed occlusion could confuse the visual features of sugarcane seedlings and increase the detection difficulty,the YOLOv8 model was able to identify sugarcane seedlings accurately. The results of sugarcane seedling identification were vectorized, and the post-processing method of spatial analysis was used to automatically rotate the crop rows to the vertical direction, and then the number, direction, row spacing, and start and end points of the crop rows were accurately obtained by means of clustering, intersection calculation and coordinate conversion, which effectively solved the problem of differences in crop row conditions between different plots and within the same plot. In 2 large plots in the study area, 8sample sample plots（A-H） with different morphology, orientation and area were randomly designated, and the seedling deficiency rates of the 8 sample plots were calculated based on the standard distribution template of sugarcane seedlings without seedling deficiency, and the results showed that the errors of the seedling deficiency detection model were4.35%, 2.98%, 4.28%, 2.91%, 1.88%, 0.51%, 1.10%, 1.51% and 1.51% respectively. In addition, based on the results of the seedling deficiency detection model, the coordinates of the location of the seedling deficiency in each sample plot could be obtained. 【Suggestion】The automated sugarcane seedling detection and seedling deficiency rate calculation method based on the YOLOv8 model can process a large amount of image data quickly and efficiently, with a high degree of automation and precision, and is suitable for detecting seedling shortage in a wide range of sugarcane fields, and can provide specific seedling deficiency coordinates. The follow-up study suggests to improve the model recall rate by multiscale detection, to reduce the problem of missed detection by using sliding window cropping image for data annotation,and to expand the dataset to improve the generalization ability and robustness of the model, so as to effectively improve the stability and accuracy of the sugarcane seedling shortage detection results.