Remote Sensing Monitoring Of Nitrogen Content In Rice Based On Information Fusion Of Multi-sensor Imagery From UAV

China is the world’s largest producer and consumer of rice,with more than 60%of the Chinese relying on rice products as their staple food.Nitrogen,the nutrient with the highest growth demand,is of great significance in the growth of rice.Therefore,precise fertilization decisions for rice fields are closely linked to the real-time and accurate monitoring of rice LNC to improve rice growth,reduce costs and protect the environment.Traditional nitrogen monitoring in rice usually requires manual field sampling,which is not only time-consuming and labor-intensive,but also subject to potential human error,and can usually only obtain information at the point level,which makes it difficult to apply on a large field scale.As remote sensing technology develops rapidly,fast and non-destructive data collection and rice nitrogen content monitoring can be achieved with the advantages of UAV remote sensing platforms and multiple types of sensors.In this study,rice in the Tianjin Demonstration Centre of The Quality Agricultural Products was used as the research object,and UAV high-definition digital and multispectral image data of rice at the jointing,booting and filling stages were collected to carry out research on remote sensing monitoring of rice nitrogen content based on the fusion of UAV multi-source image information,which can provide reference for accurate fertilization decisions in rice fields and nutrient monitoring of other crops.The main research results and highlights are as follows:(1)The study of nitrogen content monitoring in rice based on UAV multi-source image fusion and feature parameters optimization.Firstly,the UAV digital and multispectral images were fused by GS fusion method,and then the background noise of remote sensing images was removed using two HSV color space transformations combined with random forest classification algorithm.Based on the processed image data,19 typical vegetation indices were extracted as candidate spectral feature parameters,and the candidate spectral feature parameters were optimized using the successive projections algorithm(SPA)and the competitive adaptive reweighted sampling(CARS)method.Finally,two regularity machine learning algorithms,lasso regression(LASSO)and ridge regression(RIDGE),were combined to evaluate the prediction accuracy of the rice LNC monitoring model under different conditions.The results show that the fused images have both high spatial resolution and high spectral resolution,and the accuracy of the model based on fused images can be improved by7%on average.The background noise removal method can effectively separate rice,water bodies,soil and shadows in rice fields,and the accuracy of the model can be improved by 5%on average after noise removal.The RR-SPA model based on the SPA and RIDGE algorithms had the highest prediction accuracy with R~2=0.76,RMSE=10.33%and NRMSE=2.38%when fused images with background noise removed were used as the data source.(2)The study of nitrogen content monitoring in rice based on multi-type parameter fusion and deep learning.Using the fused UAV images with background noise removed as the data source,spectral feature parameters(VIs)based on digital and multispectral images,color space parameters(Colors)based on RGB,HSV and L*a*b*color spaces and texture feature parameters(Textures)based on grayscale symbiosis matrix were extracted and filtered by gray correlation.Different parameter fusion methods such as"VIs","Colors","Textures","Colors+VIs","VIs+Textures","Colors+Textures"and"Colors+VIs+Textures"are also considered and five common machine learning algorithms such as RF,GPR,PLSR,SVM and ANN are compared with a fused deep neural networks DNN-F2.The results show that color space information and texture feature information are effective in improving the prediction accuracy,and the parameter fusion based on"Colors+VIs+Textures"is the most effective,with an average improvement of 25%in R~2 and an average reduction of 8.71%in RMSE compared to the fusion methods with two parameters.The DNN-F2 model was significantly better than the traditional machine learning algorithm,with an average improvement of 33%in prediction accuracy,and achieved the best results at the jointing stage with R~2=0.72,RMSE=11.81%and NRMSE=2.72%.