Detection Of Winter Wheat Affected By Pests And Diseases And Lodging Disasters With Cost-sensitive Learning Models

Agricultural disasters are a vital role affecting the yield of crops.Early detection and effective prevention is an effective measure to reduce agricultural economic losses caused by crop pests and diseases while increase the yield and quality of grains.At the same time,it occurs to crops encounter lodging disasters during the growth process.Crops lodging could not only reduce crop yields also bring difficulties for harvesting in the later period works.The accurate area and degree of lodging estimation can provide key information for claims compensation and other links in agricultural insurance.Remote sensing technology can obtain information of crops on the growth quickly and effectively status without destroying the crops,which providing technical supports for crop pests and diseases and lodging monitoring.Nearly research about the monitoring of disease and pests and crop lodging commonly take measures of inducing disease in artificial test field and lodging simulation,but in fact,during real agricultural production process,it is also correlating between the spectral signature reflected by those diseases in field area and many other elements of environment near crops.Besides,the range of agricultural disaster are usually categorized by agronomy parameters inverted by vegetable index and then employs it to evaluate agricultural disaster.those models and methodologies just carried out in exact farming areas and regions and it is still hard to utilize in large field areas.Cost sensitive learning model which can detect crop diseases and insect pests rapidly and effectively and contribute to early prevention and measures to implement.Cost-sensitive learning bases on the mechanism of setting distinct incorrect classification cost in different application scenarios,thus improving the accuracy of classification on interesting targets and recognizing and detecting more categories in satellite images.In this dissertation,the winter wheat in Henan is taken as the research object to study the methods for monitoring wheat pests and diseases and lodging with remote sensing data.Content includes: 1)Sentinel-2A data is used as the main data source to study the sensitive characteristics of monitoring of pests and diseases in City of Xinxiang.To research the potential of cost-sensitive learning in improving the recognition accuracy of diseases and insect pests of crops,two supervised classification methods including cost-sensitive and cost-insensitive learn are taken.2)Domestic GF-2 data was used as the main data source to study the method of monitoring of winter wheat lodging in the City study area of Changge.Different combinations of feature sets are designed and XGBoost machine learning methods were used to discuss the best combination of crop lodging monitoring.The best texture feature scales were research by comparing and analyzing the result of lodging monitoring.The top findings and innovations are as follows:(1)It could be quickly and effectively to detect Winter wheat suffered pest and disease with cost-sensitive models with optimal features.1)Cost-sensitive learning methods have obvious advantages in the use of remote sensing to monitor crop pests and diseases,and can increase the recall rate of disease-related pests and diseases greatly,which is of great significance for early effective prevention.Among them,CS-SVM has increase the recall rate of winter wheat pests and diseases compared to SVM by 5%,and the overall misclassification cost has decreased by 22.The CS-Naive Bayes model has increased the recall rate of diseased and pest wheat by 5% compared to Na?ve Bayes and then reached the highest value 99%.And the overall misclassification cost lost also reaches the minimum of 26.6;2)It shows significant differences in the spectral characteristics of wheat and healthy wheat especially on the three bands of 705 nm,740nm and 783 nm bands,which were good features distinction between crop pests and health conditions.3)The feature selection method Relief has nothing to do with the specific classification model so can be used to obtain the best pest identification features.Sensitive feature combinations;4)Although cost-sensitive learning methods face the risk of taking healthy wheat into pests as diseases,it is of practical economic importance to improve the recall rate of pests by sacrificing the accuracy of some healthy wheat.(2)It could be quickly and effectively to detect winter wheat lodging based on texture and vegetable index from GF-2 data with cost-sensitive learning models.1)The cost-sensitive learning models perform better than cost-insensitive models on the monitoring wheat lodged.CS-SVM increases the recall rate of lodging wheat by 5.79% and the overall misclassification cost has decreased by 231.75.and CS-Na?ve by the 4.19%,the overall misclassification cost has decreased by 138.75;2)The high spatial resolution characteristics of GF-2 remote sensing data are very favorable for the monitoring of crop lodging,it shows significantly different on spectral reflectance of the canopy in the visible and near infrared bands.The spectral reflectance of the GF-2bands were increasing after the ripening of the wheat.Among them,near infrared band in has the largest crease rate,followed by the green band;3)The texture features with high spatial resolution are very favorable for the monitoring of lodging.Different texture feature scales have different effects on the classification accuracy and classification capabilities.The texture feature of the 5X5 pixel-window size has the highest accuracy of the lodging monitoring.The four-band mean feature has the greatest ability to discriminate the lodging wheat;4)The supervised learning model XGBoost can not only perform excellently on classification,but also can be used as feature selection;5)The combination of spectral features,vegetation index features and texture features can improve the accuracy of lodging monitoring and performed best in monitoring of lodging.