Summer Allen
WEBINAR
Using Satellite Imagery for Early Warning of Productivity Constraints
Organized by the Food Security Portal (FSP)
OCT 31, 2019 - 11:00 AM TO 12:30 PM EDT
A Press for the Planet: Journalism in the face of the Environmental Crisis
Pre-Harvest Loss Estimation in Tanzania
1. Pre-harvest loss estimation in Tanzania
Summer Allen*, Soonho Kim*, Ritvik Sahajpal**, Michael Laurence Humber**, Rachel Huang*, Samson Dejene
Aredo*, Sonja Betschart***, Rose Funja****, and Yussuf Said Yussuf***
*Markets, Trade, and Institutions Division, IFPRI
**Department of Geographical Sciences, Univ of Maryland
***Tanzania Flying Labs, WeRobotics
****Agrinfo, Tanzania
Supported by NASA Harvest and the Food Security Portal
(www.foodsecurityportal.org)
3. Satellite Data and Machine-Learning Model
• Satellite data were used to create the model and we used the UAV data to
supplement/calibrate the model
• We used the following data sources:
• Planet for hyper-local field delineation (has higher resolution for delineation of fields)
• Landsat and Sentinel-2 based HLS data for yield modeling at field scale
• MODIS data for regional and global scale modeling
• Data used: Temperature (NOAA CPC), Growing Degree Days (NOAA CPC),
Precipitation (UCBS-CHG CHIRPS), Evaporative Stress Index (USDA-NOAA), Soil
Moisture (SMAP), NDVI (Planet Doves, Sentinel-2), GCVI (Planet Doves, Sentinel-
2), UAV data collection
4. 4
Survey to ground truth the images
01
02
03
04
Total of 300 farmers
Two rounds of panel survey:
growing season/ after harvesting
GPS-based plot mapping
Drone data collection for verification
5. SenseFly eBee X Fixed-
Wing Drone + Parrot
Sequoia Multispectral
Sensor
Used for:
• Calibration for satellite data
• Identifying surface of fields
• Checking the crop condition
from the RGB camera
during data process
RGB NDVI
6. Hypothesis: End of season yield forecast is a good proxy for crop conditions.
• Crop yield forecasts (upon which crop condition assessments are based) should
have higher uncertainty at the start of the growing season.
• Each crop has a different response to the abiotic factors (temperature,
precipitation, solar radiation) and the response varies by phenological growth
stage and geography. It is important to take this variation into account.
• We assess the validity of the system by comparing mid-season yield estimates
from GEOCIF with end of season yields through back testing for previous years.
Global Earth Observations for Crop Yield Forecasting (GEOCIF)
7. How is GEOCIF Implemented?
1. Collecting Inputs 2. Preprocessing 3. Building Machine-Learning Model 4. Predict 5. Visualize
• GEOCIF predicts end of season yield based on Machine-Learning models
• End of season yield is then converted to crop condition
• GEOCIF is fully automated and is run weekly
8. How well does GEOCIF perform?
GEOCIF has a prediction error between 3-5% for key producing countries (2001 – 2016)
National level (Ukraine)Global level
9. ● NDVI is a useful indicator for differentiating
between healthy and unhealthy vegetation
(NIR≅VIS).
● NDVI tends to ‘saturate’ for healthy vegetation
when leaf canopy is dense (i.e., when NIR>>VIS).
● When analyzing healthy vegetation, GCVI could
perform better.
● Results from Burke and Lobell (2017) also
demonstrate the value of using GCVI.
GCVI may have better performance than NDVI in analyzing healthy vegetation
Source: https://midopt.com/healthy-crop/
10. GEOCIF Model
Model fit - we have built a simple and robust model for forecasting crop yields based on GCVI/NDVI:
production data from 2017 and 2018
data from above-average farm sizes (i.e., larger than 1.5 ha, depending on crop) for more
stable reading of VIs
produces crop yield forecasting after the peak GCVI/NDVI becomes available (so
approximately after half of growing season)
• Planet for hyper-local field delineation
• Landsat and Sentinel based HLS data for yield modeling at field scale
• MODIS data for regional and global scale modeling
There is a meaningful relationship between satellite data-based metrics (VIs) and observed crop
yields across maize and sunflower in the sample for Tanzania.
11. GCVI has been able to better
distinguish among the ‘Below
Average’, ‘Above Average’
performing fields when
compared to NDVI.
GCVI is especially sensitive
to dense canopies even
after NDVI saturates.
12.
13. Findings
All models are built on linking peak Vegetation index value with yields.
For white maize, using linear regression on GCVI data seems to give best
combination of R and RMSE values.
For sunflower, all models have similar R and RMSE values when using GCVI data.
When using NDVI data, linear regression model seems to work best.
2017 seems to be a harder year to distinguish between below and above average
performing fields than 2018 and may be due to recall.
The model currently does not fit well with intercropped fields but we are working
on this further and hope to expand work to Ethiopia.