| Wheat is widely planted in China and is one of the main food sources in China.Timely detection of abnormal conditions in wheat field is of great significance to ensure grain yield and safety.This paper aims to combine UAV,image processing and machine learning technology to realize the monitoring of wheat field anomalies and promote the construction of heaven earth integrated intelligent farmland.This paper analyzes three common abnormalities in wheat field,namely leaf disease,lodging and seedling loss.According to their manifestations and characteristics,different model methods are used for detection.For the experimental materials involved in this paper,the data of wheat leaves come from the network,and the data of wheat lodging and seedling loss are captured by UAV.The main research contents and results of this paper are as follows:（1）Aiming at the problem that the existing leaf disease recognition model needs to input a large number of known disease data,and when an unknown disease occurs,it will not be detected due to the lack of data,an unsupervised anomaly detection algorithm based on color moment is proposed in this paper.The algorithm first carries out image denoising,image segmentation,image interception and other preprocessing operations on the original data,and then only needs to input the normal leaves.After K-means ++ clustering and image segmentation,the color moment features are extracted,the comparison model is constructed and the threshold is set to test whether the leaves have disease abnormalities.Comparing the correct rate with the missing alarm rate,it is found that when the data is preprocessed and enough normal data sets are input,the result of this algorithm is better than traditional machine learning algorithms such as KNN,ABOD,CBLOF and so on.In the abnormal detection of wheat leaf diseases,the correct rate is 93% and the missing alarm rate is about3%.（2）In the middle and late growth stage,wheat is prone to lodging abnormalities due to the influence of its own varieties,cultivation measures,weather environment and other factors,which is one of the main factors limiting the high,stable and high quality of wheat.In order to detect the lodging anomaly in time in order to take remedial measures and reduce the loss of wheat yield,this paper uses the improved GANomaly deep learning model to detect the anomaly of wheat lodging in large area,medium area,small area and mixed area in a semi supervised learning way.Compared with many improved CAE models,the results show that the improved model,im-GANomaly,has better detection effect on wheat lodging anomaly,and the AUC value of lodging detection in different areas is not less than 90%.（3）After wheat sowing,due to lack of nutrients,improper sowing depth and soil environment,abnormal loss will occur in the emergence stage,which is unfavorable to the realization of stable and high yield of wheat.Firstly,VGG-19 network was used to identify and count the seedlings,which provided a reference for calculating the emergence rate of wheat.The average counting accuracy was more than 95%.Then the pictures and videos of missing seedlings are selected.After processing,the YOLO v5 m detection network is finally selected to realize the detection of missing seedlings in some wheat fields,and the m AP value can reach 53%.Then it realizes the visualization and information statistics of missing seedlings in wheat field composite map,which makes it convenient for relevant personnel to understand the seedling information and take corresponding replanting measures. |
