High throughput phenotyping of field crop experiments using UAVs. Ph. Burger, R. Marandel, F. Baret, G. Colombeau, A. Comar

Transcription

1 High throughput phenotyping of field crop experiments using UAVs Ph. Burger, R. Marandel, F. Baret, G. Colombeau, A. Comar PHILIPPE BURGER Drone Garden Workshop - 10/07/2018

2 Phenotyping? Genotype= the DNA code Phenotype= all the observable characteristics of a living organism (morphology, physiology, molecular, ) Flowers early (degree-day) Phenotype varies with environment Big leaves (cm²) Sensitive to mildew (1-5) High throughput plant genotyping is a reality! Yield (T/ha) and time! No big revolution in phenotyping methods (specially in field).02

3 Why high throughput? Dissecting complex traits (yield!) requires an examination of thousands of lines (Myles et al. 2009) Phenotyping is a bottleneck for plant genetics Challenge: to bridge the throughput gap between genotyping and phenotyping, specially for quantitative trait (yield, stress tolerance, ) Isolated plant: In field at canopy level: (Avion jaune, 2005) Heliaphen pot platform with an automatic watering and weighing robot Combine plant remote sensing knowhow and high resolution imagery techniques based on UAV systems and robots.03

4 Phenodrone: UAS for crop phenotyping Hexacopter 3m < altitude < 150m 3km/h < speed < 40km/h Max flying time: minutes Payload < 1kg 2 axes gimbal (camera orientation) GPS controlled flight (Mikrokopter based) Precise positioning of cameras by photogrammetry Ongoing work on RTK positionning.04

5 400 nm 700 nm 1300 nm µm 14 µm VISIBLE NEAR INFRARED MIDDLE INFRARED THERMAL INFRARED Pigment Absorption Molecule Absorption Radiation emission Chlorophyll Carotenoids Xanthophyll Brown pigments Fluorescence Flux density Silicium Water Organic compounds Protein Cellulose lignin InGaAs Temperature Emissivity Microbollometre Pixel size Cooling Price.05

6 High resolution RGB Camera Sony g (60 mm) 6000 x 4000 pixels; 21.9 X 14.6 (60 mm) 1 Hz acquisition frequency Multispectral camera Airphen 200g (+ battery) 6 spectral bands 1280 x 960 pixels, 35 X Hz acquisition frequency Autonomy: 20 min GPS/IMU recording Sensors Thermal Infra-Red Camera (triggered by Airphen camera) FLIR Tau2 TEAX (Thermokaptur) 120 g 640 X 480 pixels ; 32 X 26 (19 mm) 12 Hz acquisition frequency FFC (Flat Field Calibration) Synchronization / GPS.06

7 Photogrammetry for a precise positioning of cameras: a structure from motion approach.07

8 Flight preparation: GCPs GCP (Ground Control points) for precise image reprojection Specific geometric targets for automatic recognition Precise GPS (RTK) position Minimum 5 (better 10) targets Targets clearly visible.08

9 Phenoscript: elementary processing pipeline Step A: Data pre-processing Step B: Geo-referencing (using ground targets) Step C: Extraction (GIS).09

10 What can we measure? Crop establishment Plant density, Gaps Evaluate the quality of a field experiment, identify biased plots RGB images Crop growth Height Green fraction Leaf Area Index (m² leaf/m² soil) Plant content Chlorophyll content Nitrogen content State Water stress level VIS-Near Infra-Red images Thermal Infra-Red images.010

11 Estimation of plant density from High-Res RGB images Characteristics of flights du 09/05/2017 Altitude of flight (m) : 40 m Ground Resolution (mm): 5,5 mm Trial size : 1764 microplots (3,2 ha) Nbre of flights to cover the trial: 6 Flight duration 10 minutes Pre-treatement using PHENOSCRIPT: Extraction of images for each microplot 2578 images 26 Go (R. Marandel, D. Campergue, JF Liévain - UE Auzeville).011

12 Plant Identification Identification of rank (regression on centroïds ) Elimination of small objects away from rank (weeds) Identification of plants using sum of green pixels on each column Distance distribution between peaks (blue) & objects (yellow) Estimation of theoritical distance between plants used to detect missing plants.012

13 Results for plant densities estimates Terrain effect Effet terrain Good estimates of plant densities Map of plant densities.013

14 Use of 3D dense point cloud Digital Elevation Model to measure: heights Plant or plot volumes As long as one can pick up the ground level! soil plants.014

15 Empirical approach using vegetation indices Green fraction FV = 98.4 * MCARI2 adj.r² = MCARI Biensur Hysun Soissons Apache CapHorn Isildur N1 N2 N3 The relationship between MCARI2 and the green fraction is independent from cultivar and nitrogen modalities Need for: - additional information on structure - other directions of measurement.015

16 Mechanistic approach using radiative transfer model Reflectance depends on: Canopy architecture Sun position (zenith, azimuth) At nadir, reflectance will be very sensitive to the fraction of illuminated soil viewed Cab,Cw, SLW LAI, Leaf inclination soil parameters Sun & viewing angles PROSAIL optimisation Reflectance* Reflectance If necessary more realistic 3D model of the plant will be used Measures soil parameters Sun & viewing angles.016

17 Evaluating water stress using Thermal Infra-Red Good agreement between UAV and ground radiothermometer measurements ( Inra) Canopy temperature depends mainly on: The green fraction Contrast between soil and leaf temperature Leaf orientation Observation geometry Interpretation of TIR measurements is still difficult.017

18 What do I dream of? High precision camera positioning to reduce image processing time for geopositioning No more need for Ground Control Points!. get rid of photogrammetry! Better autonomy for faster acquisition on big trials.018

19 Combining air and ground systems on plant phenotyping platform UAV Throughput: 1000 plot/h Near instantaneous measurements of all plots Passive measurements Sensitive to light conditions Sensitive to wind Phenomobile (2017) 150 plot/h Hours between 1st and last plot mm² precision Active measurements Insensitive to light conditions Less sensitive to wind! Phenotyping Strategy: UAV frequent measurements using Phenomobile based informations on canopy structure for data analysis.019

20 Merci de votre attention.020