Parameter Retrieval Of Different Crop Types Based On Hyperspectral Multi-Angle Data Recorded From A UAV Platform

The content of physicochemical components of vegetation is an important indicator that reflects vigor and productivity of crop types.It is a key reference for the precision agriculture management,yield estimation,pest and disease detection and water nutrient stress diagnosis.The primary objective is to establish and compare different retrieval models for relative chlorophyll content of various crop types based on multi-angle observations and estimate the impact factors on the accuracy of measurements.In a first step,the acquisition-,calibration-,and smoothing methods of UAV recorded hyperspectral data were studied and analyzed.Several retrieval models to estimate the relative chlorophyll content were established and compared for five different crop types(cabbage,carrot,green onion,white radish,cabbage).Validated by in situ measurements,the most suitable retrieval model was obtained and applied to the entire hyperspectral imagery.In a second step,the bi-directional reflection distribution function(BRDF)of four crop canopies(carrot,green onion,white radish,cabbage)was quantified and assessed to find the optimum angle for most detailed information extraction.The main achievements and conclusions are as follows:(1)The image pre-processing process flow and the methods of UAV hyperspectral data acquisition and processing were described,and a suitable radiometric calibration function for the used device was obtained.Flat field methods has been selected after comparing three different atmospheric correction methods.A SNR based piecewise smoothing method was proposed to preserve the absorption and emission features of vegetation after application to the data.(2)Resting upon the image data collected by the UAV hyperspectral system and the corresponding measured chlorophyll content in the field,three types of models to assess the relative chlorophyll content were established.Models are applied to single crop types and to all crop types simultaneously.One sort of model used belong to the univariate and multivariate type and they were associated with selected different vegetation indices.A second type of model applied was a univariate regression model dedicated on two specific characteristics of vegetation canopies.The third model type is a partial least squares regression model based on the sensitive bands of different crops.It turned out that among all models the multivariate regression model based on vegetation indices is ideal for retrieval of most crop types.Only for the onion,the univariate logarithmic model based on the single index OSAVI shows a better performance.The results from the model type based on the red edge spectral feature are acceptable,but the performance of this model is not as good as the two approaches mentioned above.The results of the PLSR model turned out not to be ideal for single crop types,whereas it shows a higher R-square for the prediction of all crops.The optimal retrieval model suitable as well for each single crop as for all crops was established and applied to the entire hyperspectral imagery.This way,a distribution map for the relative chlorophyll content of the study area was obtained.The result shows that it is infeasible to establish one quantitative retrieval model of relative chlorophyll content based on statistical methods that is suitable for all vegetation types.(3)Furthermore,the spectral characteristics of all crop canopies under consideration have been investigated and assessed according to different observation angles(elevation and azimuth)and an optimum information extraction.The overall reflectance of vegetation canopies varies with the insolation-,and the observation angle.The gradient of increase resp.decrease of reflectance at different viewing angles can be linear or non-linear at any wavelength.In the near-infrared range,the gradient of reflectance is relatively stable.In the visible range,the gradient usually increased along with the observation angle,due to the influence of the soil fraction and shadowing of the respective crop type.For different solar elevation angles,the reflectance of cabbage and radish has linear relationship with solar elevation angles in the near-infrared range.But that of onions and carrots exhibit a non-linear and non-monotonic response.Besides,in the visible range,the relationship between reflectance and solar elevation angle shows a lower regularity.Concerning observations in off-nadir mode,the sensor covers more leaf area(chlorophyll)and less shadows or soils per pixel.Thus,the signal to noise ratio is higher,which results in an increased spectral/radiometric resolution and finally in a more accurate information retrieval.(4)The optimal chlorophyll retrieval model was used to estimate the relative chlorophyll content from the average spectrum obtained from different insolation??and observation angles,and the effects on the model retrieval results were analyzed.The interference percentage of different observation angles(200 interval)on the chlorophyll content retrieved of cabbage(pure pixels),radish and onion is about 10%,but the interference percentage on the retrieval model of carrot reaches 41.7%.For different solar elevation angles(6~7° interval),the interference percentage to the retrieval model of radish and onion is about 3.7%,and the interference percentage to the retrieval model of cabbage(mixed pixel)is 13.4%.The level of interference with the carrot retrieval model reaches 43.3%,so the carrot retrieval model itself may not be robust enough that is interfered seriously by the observation and solar angles.The comprised results above clearly reveal the necessity for further developments of multi angle analyses and related methods,as nearly all future optical Earth observing satellites provide multi-angle observations to increase their revisit time.In this context,studies to optimal observation angles for an improved information extraction are crucial for parameter retrieval in agriculture monitoring and analyses.