UAV-based Hyperspectral Remote Sensing System

Hyperspectral technology is a new remote sensing technology rapidly developed at the end of the 20th century,which is capable of acquiring the target image and the spectral data of each pixel point on the image simultaneously in the ultraviolet,visible and infrared regions.By measuring the changes in spectral reflectance and its related characteristic parameters caused by vegetation pest and disease infection,large-scale plant pest and disease trends,condition monitoring and early warning can be performed.Pine wilt disease is the most serious international forestry quarantine object affecting our forest ecology,and is also a destructive forest disease that is closely guarded by countries around the world,with wide spread,fast onset and high mortality rate.Early detection and treatment is of great importance to reduce the economic losses and ecological damage caused by pine wilt disease.In this study,we developed a new remote sensing device with a multi-rotor UAV equipped with a hyperspectral camera based on liquid crystal tunable filter to collect,process and analyze the information on the change of the reflectance spectral characteristics of the diseased vegetation,analyze the correlation between the spectral curve and the disease index using a support vector machine,and determine the location of the diseased plants by combining with a deep convolutional neural network.The aim of this study is to locate the location of the diseased plants.The aim is to provide a research basis for an effective and accurate new technology for locating the latent trees of pine wilt and for disease monitoring and early warning,and to provide the conditions and basis for containing and eradicating the continuous spread of pine wilt in China.The research contents and conclusions of this paper are as follows.1.A leaf surface spectral acquisition device based on a line array CCD grating spectrometer was developed.The device adopts an adjustable optical path scheme to simulate the effects of different lighting conditions on the reflected spectra of leaf surfaces;a rapidly replaceable sample plate and base were designed to improve the spectral data acquisition efficiency.The visible-near-infrared(NIR)spectra of the pine needles under different disease stages of pine wilt were measured,and the spectral reflectance,first-order differential spectra and disease stage index were correlated.From the results,the effects of early pine wilt stress on the original and first-order differential spectra were mainly concentrated in the red-edge band of680-740 nm,and the intensity of the reflectance spectra in the band was strongly influenced by the change of disease course,which could reflect the development of pine wilt disease course.The parameter extraction of the reflectance spectra of the pine needles was performed,and the correlation analysis of the disease course indices was carried out using conventional vegetation indices such as NDVI and RVI.The normalized difference moisture-disease course index(NDWDI)based on the sensitive bands 698 nm and 720 nm was proposed.A prediction model was developed for the spectral data and disease course index using a support vector machine with kernel functions including linear,quadratic polynomial,cubic polynomial and Gaussian functions.The results showed that the prediction of pine nematode disease course index by the support vector machine using multi-band co-computation improved to 80% accuracy compared to single-band linear and nonlinear regression.The range of multi-band selection was thus determined to be 691,692,709,710,715 and 718 nm.2.A hyperspectral camera based on LCTF liquid crystal tunable filter was developed and mounted on a multi-rotor UAV to take hyperspectral images of the ground vegetation canopy of the Horsetail pine woodland.The multi-rotor UAV is controlled by pixhawk flight control system,and uses Mission Planner as a ground station to monitor the flight status and generate a shooting route trajectory to the target forest area using automatic planning.The hyperspectral camera is equipped with an stm32 chip as the controller,which can control the transmission wavelength of the LCD adjustable filter by modulating the square wave signal and control the hyperspectral camera wirelessly during the flight operation.The hyperspectral images are segmented using a sliding window image preprocessing method,and the individual segmented images are labeled to form an aerial photography dataset of horsetail pine forests.A deep convolutional neural network with Inception-Resnet structure was established to perform feature extraction on the dataset.The results show that the model can reach 85% accuracy in judging the late stage of disease,which has higher accuracy compared with the traditional random forest model.An image comparison model based on normalized difference moisturedisease course index was established to combine the hyperspectral images of the canopy of Marestail pine for disease course judgment with an accuracy of 73%.