Development And Application Of A Multispectral Remote Sensing System Based On The Multi-rotor Drone

The production of crops in China depends on the application of chemical fertilizers.There are problems such as low fertilizer efficiency and serious waste loss.In this paper,aiming at the current situation of variable decision-making for low fertilizer application in China,a miniature multispectral snapshot imaging system with small size,light weight,and high imaging speed has been developed based on a low-altitude remote sensing platform.The rice nutrient information was obtained from the individual plant canopy scale and field scale,and the SPAD value was used as the indicator of nitrogen assessment.Remote sensing inversion models were established to obtain the visualized spatial distribution map,which was used as a reference for the generation of the variable fertilizer prescription map at the regional scale.The main research work and achievement were as follows:(1)Based on a miniature multispectral snapshot imaging sensor,a multispectral imaging system with small size,light weight and high imaging speed was developed.According to the characteristics of the low-altitude remote sensing platform,a multispectral imaging remote sensing system was designed and integrated based on the CMOSIS CMV2000 multispectral imaging sensor provided by IMEC.The system consists of five parts:the computer,multispectral camera,data storage,interface display and power supply system.It realized the power supply,control and image data storage of multispectral sensors,and enabled fixed-point capture according to the real-time requirements of remote sensing operations.Dark current of the multispectral camera and the spectral and spatial information of rice canopies were tested.It was initially verified that the system can be used for the acquisition of rice nutrient information.(2)Based on the low-altitude simulation platform of UAV,the SPAD values of rice in single canopy scale were detected.A UAV low-altitude simulation platform equipped with the multispectral imaging system was used to acquire the multispectral images of rice canopies.The PC2 image was selected as the feature image for threshold segmentation to achieve automatic recognition and extraction of target regions.A standard four-gradient calibration plate was used to convert the DN values of samples to surface reflectance using the empirical linear method.The prediction models of SPAD values of the rice canopies were established using PLS,BPNN,LS-SVM and ELM algorithm based on the full-band spectral information.And the accuracy of the ELM model was better than that of the other three models.The Rc of the modeling set and the Rp of the prediction set were the highest of all,which were 0.93 and 0.84 respectively.Based on the combination of the three narrow band vegetation indices,the optimal vegetation indices were used to establish the prediction models of the rice canopy SPAD values.It was found that the modeling effects based on the NDVI-like exponents of 693 nm and 849 nm were slightly better than those in the other two indices.The correlation coefficient of the prediction set was higher than 0.75,and the RMSE of the prediction set was less than 1.6.(3)Based on the eight-rotor UAV,the SPAD values of rice in regional scale were detected.Prior to the acquisition of multispectral images,suitable test parameters were selected according to the requirements of remote sensing tasks.Agisoft Photoscan was used to achieve rapid and automatic mosaicking of UAV multispectral images,and the UAV multispectral orthophotograph was obtained.The combination of site alternative calibration method and empirical linear method was used for radiometric calibration to obtain the average reflectance curves of rice canopies in field under different nitrogen gradients.The PLS algorithm was used to establish the prediction model of SPAD values of rice in field based on full spectra.The R2 of modeling set was 0.8135,and the R2 of the prediction set was 0.6759.A linear prediction model for SPAD values of rice in the field was established based on the combination of three narrowband optimal vegetation indices.It was found that the prediction model for SPAD values of rice in the field based on SR-like(R732,R889)had the highest accuracy.The R2 of modeling set was 0.7008,and the R2 of the prediction set was 0.6354,although the accuracy was not higher than the PLS model based on full spectrua,it simplified the calculation.Based on the linear model of SR-like(R732,R889),the multispectral images were visually inverted and the spatial distribution of the visualization was realized.