New Facts and Follies of remote sensing of tree crops · 2019. 6. 10. · visual assessment from...
Transcript of New Facts and Follies of remote sensing of tree crops · 2019. 6. 10. · visual assessment from...
Facts and Follies of remote sensing of tree crops
Professor Andrew [email protected]
Not all about the Tech
Cost/ benefitof adoption
Linking with core disciplines
Who do you Trust?
Address relevant constraints
Appropriate calibration,validation
The Tech
What The Farming ?
WHAT is Remote sensing:
Defined as the acquisition of information about an object or phenomenon without making physical contact…
Sensor Resolution
Temporal Spectral
SpatialRadiometric
Frequency of data collection
Number and width of spectral bands
(Multispectral/ Hyperspectral)
Pixel size and area covered (swath)
Data sensitivity (bits)8 bits (28)- 256 level Landsat11 bit (211)- 2048 levels IKONOS16 bits (216)- 10000 levels WV3 SWIR
Spectral characteristics of a leaf
http://www.innovativegis.com/basis/pfprimer/topic7/Topic7-8.gif
NDVI
NDVI: is a good indicator of tree vigour (size and health).However, other Vis may be better correlated to the parameters you are interested in.
Examples of structural and pigment based vegetation indices
Hyperspectral (Spectral Resolution)
http://www.markelowitz.com/Hyperspectral.html
It is the Holy Grail?
IKONOS 0.8m resolution
Sentinel 10m resolution
Cane
Macadamia trees
0.8 m image: $24 km2 10 m image: $0.00km2
Spatial Resolution
Remote Sensing Technologies
Satellite Field based Sensors
Tractor based Sensors Airborne SensorsUAV/ Drone
https://earthzine.org/wp-content/uploads/2012/08/CEODA-RD-Cover.jpg
Satellite
Comparison of relevant satellite platforms for Ag.
WV-3 Sentinel-2 Geo-Eye -1 RapidEye Spot 6-7Landsat 8-
OLI
Revisit time > 5 days 5 days > 3 days 1-6 days As required 16 days
Pancromatic
Resolution0.31m N/A 0.46m N/A 1.5m 15m
Multispectral
Resolution 1.24m
10m, 20m
and 60m 1.84m 5m 6m 30m
Number of
Multispectral
Bands
8 (16 with
SWIR) 13 4 5 4 10
Cost ~$45/km2 Free ~$30/km2 ~$1.60/km2 ~$1.80/km2 Free
Satellite Sensor
~$5.50/ha ~$3.00/ha ~$0.16/ha ~$0.18/ha
Drones, UAS, UAVS, RPAS etc.
Fixed WingQuadcopters Octocopters
Blimps and balloons HelicoptersOther????
Benefits:- Can be deployed at short notice (between cloud)- Relatively low cost platform- Can provide ‘over head’ imagery of
agricultural crops in varying forms(RGB, NIR, Thermal, Lidar).
- Optical height models
Limitations:- Generally uncalibrated imagery- Limited flight time- Line- of- sight restrictions- Privacy issues- Require certification to operate commercially.- On-going maintenance and repair costs.- Data processing and interpretation
Field work – RPAS campaign
27/10/2017 Research Project Mid-term Review 14
Ground control marker
Mission planningDrone and camera
Reflection targets
28 flight configurations:
• Above ground altitude
• Image side overlap
• Flight speed
• Flight time
• Flight pattern
100 m75 m50 m25 m
60%70%80%90%
3 m/s5 m/s8 m/s10 m/s
>40°40 °30 °<20 °
Optimal Settings for Remotely Piloted Aircraft Systems Applications in HorticultureYu-Hsuan Tu Supervisors: Stuart Phinn, Kasper Johansen, and Andrew Robson
• Flying along the tree row improves the accuracy
• Flying at high solar elevation
• Flying with a gimbal improves the accuracy
• Flying with high forward overlap improves the accuracy (80%)
• Higher image resolution improves the accuracy
• recommended flying altitude is around 75 m AGL to obtain a 2 cm
• Recommended image sidelap is around 75%
• Ground calibration panes for radiometric and geometric correction
Yu-Hsuan Tu : UQ and UNE PhD student
UAV camera systems
https://www.micasense.com/rededge/http://diydrones.com/profiles/blogs/sequoia-in-the-wild
Parrot SequoiaMicro Sense
Blue: 475nm x 20nm,Green: 560nm x 20nm, Red: 668nm x 10nm, Red-Edge: 717nm x 10nm, Near Infrared: 840nm x 40nm.
Green: 500nm x 40nm, Red: 660nm x 40nm, Red-Edge: 735nm x 10nm,
Near Infrared: 790nm x 40nm.
Slant Range
Green: 550nm x 40nm width, Red: 650nm x 40nm width, Red-Edge: 740nm x 10nm width,
Near Infrared: 850nm x 100nm width.
http://www.slantrange.com/3p/
Blue: 446nm x 60nm width Green: 548nm x 45nm width Red: 650nm x 70nm width Red Edge: 720 nm x 40nm width Near-Infrared (NIR): 840nm x 20nm width
https://sentera.com/product-category/sensors/
Sentera
2016-17, 2 fields, 4 varieties, 3 sensors
Micro sense
Comparable Out puts: Foliar Nitrogen measurement in Rice
Is 2 cm imagery the best?
2 cm resolution good for measuring individual canopies
Can loose spatial integrity of the orchard
Reduce spatial resolution to get ‘zones’
How you will use the data is an important consideration
3D modelling of tree growth factors
P.J. Zarco-Tejada et al (2014)
RMSE 33 cm
P.J. Zarco-Tejada et al (2014)
Rahman et al (2018)
Thermal
Figure 3. Thermal image filtering; (a) original infrared thermography, (b) corresponding
temperature frequency distribution of the original image, (c) infrared thermography after
filtering, (d) corresponding temperature frequency distribution after filtering.
https://www.nnbusinessview.com/news/drone-technology-with-deep-roots-in-northern-nevada-
primed-for-the-future/
Salgadoe et al (2019)
Need a wet and dry reference target
Mathematical approach
Serving the needs of farmers since 2002
Industry Examples: Integration of
technologies
Mapping block yield
False colour WV image with sample locations
Area: 11.3 ha (9*6 spacing)Ave. 3.7 T/ha (20 kg/ tree)Total: 42 tonnes
Linear relationship between measured yield and corresponding canopy spectral data (VI)
Extraction of canopy data for all trees
Calculation of average/ total block yield and derivation yield
map
Satellite Imagery: Identifying spatial and temporal variability in tree vigour, yield, fruit size, phytophthora, nutrition etc.
Classified NDVI of orchard block (Worldview imagery).
High NDVI Mid NDVI Low NDVI
Improvement on current methods
Comparison between actual yield (t/ ha) to that predicted from satellite imagery and that made by visual assessment from growers. Note that no growers estimate was supplied for Block 5.
Yield forecasting accuracies for Mango
Comparison of actual yield (kg) to that predicted from WV3 imagery for 2016 -17 and 2017 – 18 seasons for three mango blocks in Northern Territory.
Predict loss
High Medium Low
TREE VIGOUR CLASS
Control Treatment
One of classified NDVI Avocado blocks surveyed in this study
Increasing yield via targeted pollination
- 5 blocks/ 18 trees per block- 4 reps of low, medium and high vigour received additional
hand pollination. - 2 reps of each class retained as controls.
R G B
ave temp 24.7 24.1 23.9
max temp 34.5 32.6 31.4
min temp 18.1 18.3 18.4
ave RH 78.8 80.5 80.9
max RH 100.3 100.0 99.7
min RH 46.9 50.5 53.2
Temperature data
Low NDVI trees higher max temp (~3+degC), and lower average relative
humidity than the High NDVI trees.
Block 2 R G B
ave temp 24.2 23.8 23.6
max temp 33.0 31.4 30.1
min temp 18.6 18.8 19.1
ave RH 78.6 79.2 80.7
max RH 99.3 98.4 97.8
min RH 47.2 50.2 54.9
(20 Nov 2015 to 28 May 2016)
12 sensors per block4- low NDVI4- mid NDVI4- high NDVI
Avocado
Mac
Possibility for thermal time model
development for Avocado maturity
Linking RS with Sensor Networks• High density sensor networks across our farms
Sensor Network o What density is actually needed?
o Sensors cost money to purchase and then to maintain.
o Data collection, QA’ing and analysis?
o Would a lower density be sufficient
o What information do you actually need?
Example Peanut industry: Aflatoxin warningAflatoxin risk is high when soil temperatures reach 25ᵒC to 32ᵒC and low soil moisture. Strategically located sensors (within farm and across regions) linked with an App warning system can assist with improved management to reduce risk/ harvest segregation etc.
National mapping of commercial orchards
Regional Mapping
Young
trees
Serving the needs of farmers since 2002
Cyclone Debbie
BIOSECURITY
Yield TCH
2018 sugarcane yield forecasting accuracies from both SPOT and Landsat algorithms (12 Australian growing regions). Algorithms also applied to derive image products for ~100,000 individual crops annually.
102
82 83
116
6773
82
68
105
145
90
127
101
149
86.786
68
126
85
92
59
100
130
86
124
98
132
86 8683
124
73 75
91
70
116
166
102
127
118
144
83.5
0
20
40
60
80
100
120
140
160
180
Region 1 Region 2 Region 3 Region 4 Region 5 Region 6 Region 7 Region 8 Region 9(1yr)
Region 9(2yr)
Region10 (1yr)
Region10 (2yr)
Region11 (1yr)
Region11 (2yr)
Region12
Ave
rage
Re
gio
nal
Yie
ld (
TCH
)
Sugar Growing Region
Actual ave regional Yield (TCH)
Pred. SPOT Yield estimate (TCH)
Pred. Landsat 'time series' Yield estimate (TCH)
Machine Learning
Example: Onesoil:5 years development1 million users
20 crop types in Florida Peanut crops in Florida
USDA:64,750 ha of Peanut
~6% accuracy
Summary
• Remote sensing is an effective ‘tool’ for identifying the spatial and temporal variability in tree health and production.
• It needs the appropriate calibration and validation when used to predict specific variables.
• It is a ‘tool’ to help with core sciences (agronomy, pathology, breeding etc). It is not a silver bullet.
• Just cos it looks pretty doesn’t make it right
Where to for Citrus
• Currently doing a small pilot study with WA citrus and Arbocarbon
• Submitted a tender to Hort. Inn. evaluating the best platforms and analytics to detect host trees in NT urban areas.
• ‘Rural Research and Development for Project’ submitted with decision this month.
Launched next month
Objectives
Coordinate the delivery of remote sensing capabilities by approved providers, including
researchers, public and commercial services.
Facilitate greater cross-sectoral collaboration and shared access to expertise and infrastructure.
Accelerate innovations in remote sensing technologies and their translation into on-farm
applications.
Deliver education and extension programs to enable the adoption of farming informatics and
sensing technologies, as well as help develop demand for these industry capabilities.
UNE Precision Agriculture and Remote sensing
Professor Andrew [email protected]