A Research Of The Stitching,radiometric Calibration Of UAV Snapshot Hyperspectral Images And The Applications In Wheat Nitrogen Monitoring

Hyperspectral remote sensing is an important way to quickly and non-invasively obtain crop physiological and biochemical information in precision agriculture.Currently,with the development of UAV and miniature imaging spectrometer,UAV hyperspectral imaging systems have been promoted in some research and application areas.UAV hyperspectral imaging systems have obvious advantages over traditional satellite,aerial and proximal devices for hyperspectral information acquisition.It has obvious advantages with flexible data acquisition,low cost,and high efficiency,and is suitable for hyperspectral information acquisition in small and medium area.Leaf nitrogen concentration(LNC),which is closely related to crop growth health status,is an important index for the diagnosis of crop nitrogen nutrition and is also an important reference standard for guiding nitrogen fertilizer applications in the agriculture sector.Applications of UAV hyperspectral platform can significantly improve the monitoring efficiency and capacity of crop nitrogen nutrition status.Currently,the mostly used pushbroom hyperspectral remote sensing images are large in data size and hard to acquire and process,which are the main barriers for their promotions.Unmanned Aerial Vehicle(UAV)snapshot hyperspectral remote sensing system is one kind of new hyperspectral images acquisition system.The stitching and precise calibration are the prerequisites for applying snapshot hyperspectral images.However,few existing researches systematically study the stitching and calibration of snapshot hyperspectral images.This paper aimed at obtaining hyperspectral DOM with high geometric and radiometric precision and put the DOM into the inversion of wheat leaf nitrogen content to test their capacities in practical applications.This paper chose the winter wheat at National Experiment Station for Precision Agriculture in Xiaotangshan area,Changping,Beijing as the research target.The snapshot hyperspectral images were acquired on 26 April 2015,and the leaf nitrogen content values of 48 sampling plots were sampled and measured synchronously with the flight of UAV.Also,the hyperspectral images of HG-1 calibration light source and Labsphere integral sphere were obtained in the laboratory.The main research contents and achievements include:The stitching of hyperspectral images.To get a hyperspectral DOM with geometric information,firstly we analyzed the characteristics of the POS of UAV and proposed a method to interpolate the POS data of hyperspectral images to get their position and orientation information.Also,the accuracy of interpolated POS was validated.Then we introduced a method to stitch the low-resolution hyperspectral cubes using the 3D mesh model built with hyperspectral panchromatic(PAN)images.The results showed that:a)The generated hyperspectral DOM using the interpolated POS and 3D mesh model built by PAN images had a high relative geometric precision.The difference in the horizontal direction was lower than 0.05 m(RMSE=0.035)compared with a DOM generated from digital camera images.b)The generated DEM can be used for the extraction of crop heights,and the determination coefficient was 0.680(RMSE=0.069)compared with ground-measured values.The calibration of hyperspectral DOM.First,we conducted spectral calibration for the hyperspectral sensor and evaluated its radiometric response linearity and variation.Then we proposed a multi-target-based radiometric calibration method and tested its precision.The results showed that:a)the hyperspectral sensor had a high radiometric response linearity which exceeds 0.998 for all bands.The proposed method for correcting radiometric response variation achieved a good result with the coefficients of variation were lower than 0.01 for all bands after correction.b)The multi-target-based radiometric calibration method had a good calibration result with the differences were lower than 5%for all bands compared with ground-measured values.The inversion of wheat LNC based on hyperspectral DOM.We adopted two type of spectral indices,namely the NDSI and RSI,and BP neural network methods to inverse the LNC and mapped the thematic map of LNC distribution.Results indicated that:Both the spectral index and BP neural network methods had a good inversion accuracy.NDSI achieved a better result than RSI,with R~2 of 0.605(RMSE=0.303);The modeling R~2 and RMSE of BP neural network reached 0.999 and 0.0092,while R~2 and RMSE of verification were 0.745 and 0.388,respectively.