16:00 - 16:30
The use of UAVs for high-throughput field phenotyping applications
French National Institute for Agricultural Research | INRA
Phenotyping consists in characterizing the structural and functional traits for helping breeders to select the genotypes the better adapted to each environment. The recent technological advances have made available new non-invasive methods allowing to monitor with high-throughput large panels of genotypes. Multicopters that may carry for 10-20 minutes a payload of few hundred grams up to few kilograms at variable altitude and speed are very well suited for field phenotyping. A range of sensors may be used including high-resolution RGB cameras, multispectral and hyperspectral cameras, thermal infrared cameras and LiDARs. Further, low altitude flights provide spatial resolution better than a fraction of mm to identify small organs or disease symptoms. A review of current accessible phenotyping traits is presented from such UAV observations. Several examples are given to illustrate the versatility of UAVs for high throughput phenotyping. The interpretation methods need to be well suited to provide estimates of canopy traits decontaminated from possible artifacts due to the variability in the environmental conditions during the measurement or due to confounding effects. Conclusions are drawn on the development and prospects of such UAV observations for high-throughput field phenotyping.
Multi-scale Observations For An Improved Detection Of Plant Diseases By Hyperspectral Imaging
1University of Bonn, INRES - Pflanzenenkrankheiten & Pflanzenschutz, Germany; 2Insitut für Zuckerrübenforschung (IFZ), Göttingen, Germany
Plant diseases are a highly relevant target for remote and proximal sensing. The interaction among a host and the specific pathogen results in specific symptoms which can be detected and identified by suitable sensor techniques. The identification of plant diseases on the leaf scale based on color images is subject to current research, mainly by deep learning methods in combination with smartphones. An early differentiation by hyperspectral imaging is applied on different scales. On the field scale first approaches emerge on a scientific basis. Given a suitable interpretation model, accurate detection, identification and quantification is possible even at a very early point in symptom development.
However, there is a discrepancy between optimal sensing and application scale. Exemplary, the spectral characteristics of symptoms of plant diseases are captured with a high accuracy in the laboratory under controlled environmental conditions. Sensing of the same symptoms in the field is much more complicated due to plant geometry, sensing distance and illumination conditions. Under such conditions the generation of a labeled set of high quality training data is nearly unfeasible.
This gap may be bridged by applying the training data from a more controlled measurement scale to derive a model that can be used to draw conclusion also under challenging conditions. Such a model has to compensate specific effects, e.g. different geometric effects in combination with diffused and directed sunlight.
In this study dealing with hyperspectral images in the visible (400-700nm) and near infrared (700-1000nm) range leaf rust on wheat was investigated. Training data on the leaf scale was used to derive a model transferred to the plant scale. Angular and distance related effects are compensated by a Standard Normal Variate (SNV) approach. It calculates a new data representation excluding additive and scale distortions. A sparse representation approach was selected to consider the increased mixed pixel effects due to blurring, reduced spatial resolution and high leaf angles. This approach performs a matrix factorization extracting the composition, meaning the ratio of symptom characteristics of each pixel.
We investigate the effects of SNV on the prediction results, also compared to the more established vegetation indices. Time series observations of infected and healthy wheat plants were used to validate the approach. Based on the potential to quantify the level of infection, an evaluation of the different approaches was performed.
Building on these investigations, we propose the idea of a multi-scale interpretation model for hyperspectral images. Such a model will realize an information flow from the single spore up to the satellite level spanning the gap between the optimal detection and application scales of phenomena like pathogenesis, in-field distribution and epidemiology.
Sensor Fusion As Tool For Estimating Forage Biomass In Heterogeneous Pastures
Grünlandwissenschaft und Nachwachsende Rohstoffe, Universität Kassel, Deutschland
Feeding of livestock on pastures requires constant monitoring of forage quantity to ensure consistent levels of animal and milk production. Ground based remote sensing technologies have been recognized as practical means to estimate various vegetation parameter at the ﬁeld scale. Real-time mobile sensors, which allow the collection of geographically referenced data, have proven to be useful for in-ﬁeld monitoring of vegetation characteristics such as biomass with high spatial resolution for large areas. Mobile automated sensor measurements can provide high sampling density at a relatively low cost to generate maps representing both spatial and temporal variation of the respective vegetation characteristic. As the quality of the vegetation parameter changes over a growing period, single sensor use for the parameter estimation might not be enough. Here, mobile devices have an additional value, as they can carry different sensors, which can in combined use improve prediction accuracies of vegetation characteristics. Particularly, sensor fusion of e.g. spectral sensors and ultrasonic sensors have been shown an improvement for the prediction of vegetation biomass in heterogenous pastures.
The aim of this study was, to evaluate the applicability of ultrasonic and hyperspectral mobile measurements fusion to map spatial and temporal yield variation in pastures with different grazing intensities.
Data were sampled from a long-term pasture experiment in central Germany. Three pastures with different levels of grazing intensity (moderate, lenient, very lenient) were selected. Within each pasture a 30 m×50 m study plot was established. Field measurements were conducted at four sampling dates in 2014 for each study plot. The three paddocks we scanned using an electrically driven cycle-based four-wheel-vehicle with a track gauge of 180 cm Three ultrasonic sensors and one HandySpec spectrometer (400-1600nm) were mounted at the vehicle. The exact location of the mobile system was obtained using a differential GPS system. For calibrating the sensory model, biomass was clipped on 18 reference plots per study plot and sampling date.
The prediction accuracy for forage biomass using information from both sensor was good (R2 = 0.72) and improved the sensor calibration models based only on one sensor technique. Using variogram analysis the model was applied on ~2000 mobile measurement points covering the whole study plot and the forage biomass was interpolated to the full area. The variogram analysis revealed a decrease of the range parameter from moderate to very lenient grazing intensity, suggesting a larger small-scale variation of forage biomass in the grasslands, which corresponds to the expectations.
The results of this study indicate, that the application of novel remote sensing sensors in forage science can improve the understanding of variation in forage quantity in pastures. This knowledge can help to make grazing practices more efficient.
Using UAVs for high-throughput phenotyping of wheat in heat- and drought stress environments
Techniche Universität München, Deutschland
Wheat is one of the most important cereal crops in the world. However, wheat production is facing major challenges. The growing world population with an expected size of more than nine billion people by 2050 leads to an increased demand for agricultural products. At the same time, the area of arable land for food production decreases worldwide. Therefore higher yields per hectare to assure food availability are needed all over the world. However, yield is limited by external factors such as unfavorable environmental conditions. Changes in climate, the increasing frequency of extreme weather conditions, and the increase of heat and drought events affect the agricultural production with the higher abundance of plant physiological stress. In order to ensure food security, it is necessary to breed wheat varieties that can withstand extreme hot and dry conditions. Finding suitable varieties, which could be used for selection and breeding for future climate scenarios, needs detailed information about physiological changes due to abiotic stress factors.
Different hand-held devices, such as spectrometers and thermal cameras, can be used to record the required data. Depending on the size of the experimental site, however, this can be quite time-consuming. During the last years, the use of unmanned aerial vehicles (UAVs) showed promising results trying to find a solution for high-throughput phenotyping of plant physiological characteristics. UAVs can be equipped with different RGB- and thermal cameras, as well as with hyper-spectral sensors. These allow determining not only the actual plant temperature, but also to assess spectral reflectance of the vegetation, which can be used gathering information about the health and development of crops. Advantages of the UAVs are the more flexible application compared to satellites and the rapid acquisition of many data compared to hand-held gadgets. Information about water-status of the plants, nutrient uptake and content and grain yield prediction can be assessed in a non-destructive way during the plant development.
In a field trial taking place in Moldova 40 wheat varieties were tested, 20 of which are from Germany, 20 from Eastern European countries. The latter are better adapted to the continental conditions, which can be found in the northern part of Moldova, characterized by hot and dry summers with little precipitation. During the growing season 2016/2017 regular drone flights were conducted over the site using RGB and thermal imaging in order to determine plant physiological statuses.
Using UAVs differences in senescence between German and Eastern European varieties could be detected. German varieties stay green for a longer time, as they are bred for a longer vegetation period, and tend to have premature ripening to ensure progeny, when environmental stress gets too strong. Varieties from Eastern Europe start senescence earlier, showing less signs of drought stress and more stable yields. Preliminary results also showed that groundcover and plant temperature can be determined reliably in high-throughput using UAVs.