IFPRI-Using new technologies to validating Crop-Cutting Experiments-Michael Mann
1 like716 views
The document discusses using new technologies like satellites, drones, mobile devices, and machine learning to improve the accuracy of crop yield estimates and validation of crop cutting experiments. It outlines how various remote sensing platforms and sensors could be used to collect high-resolution plant characteristic data on factors like plant height, density, and health. Computer vision techniques on cell phone imagery could also count grains and flag disease. The data collected could help address sources of bias in traditional crop cuts and potentially generate more accurate yield estimates and loss assessments through machine learning models trained on the diverse data sources.
IFPRI-Using new technologies to validating Crop-Cutting Experiments-Michael Mann
- 1. USING NEW TECHNOLOGIES TO VALIDATING CROP CUTTING EXPERIMENTS Prof. Michael Mann Dept. Geography George Washington U michaelmann.i234.me/wordpress
- 2. Calculating Actual Yields • Conceptually very simple • Crop cuts consistently found to be as biased as farmer predictions Percentage error by type
- 3. Satellite Platforms New moderate to high resolution satellite data are available: • New microsatellites 3-5m resolution • New 30m resolution satellites • Collect wide range of properties • Visible light • Infrared • Thermal Micro Satellites 3-5m resolution, Daily
- 4. Drone/Plane Platforms Drones powerful, flexible, but expensive and computationally intensive • Provide platform for specialty instruments • Thermal • Infrared • Lidar • Costs/computation time decreasing rapidly
- 5. Mobile Platforms New low cost tools can be used to collect data on: • Plant health • Cell phone imagery • Plot data through mini surveys • Planting/Harvest dates • Input use • Weather • Disease/pests • Farmer directly linked • More efficient random sampling possible
- 6. Sources of Crop Cut Bias There are a variety of sources of bias introduced into crop cuts by enumerators Measurement problems 1. Inappropriate use of tools (scales, poor records etc) 2. Failure to account for disease that make unharvestable 3. Non-random or biased location of test plots • Enumerators avoid low-yield areas 4. Failure to account for ripening or harvest over time 5. Lack of accountability for enumerators
- 7. Problem: Measurement Error Plant Characteristics New (and cheap) ground based LIDAR can quickly estimate: • Row spacing • Plant density • Plant height / biomass • Lodging • Sowing method Importantly measurements could be taken rapidly in multiple locations in field
- 8. Ground based LIDAR examples:
- 9. Problem: Measurement Error Head properties After hand threshing cell phone cameras and machine learning can be use to: • Flag potential disease / damage • Count grains • Count heads • Crop stage • Flowering/ripening etc
- 10. Problem: Timing of crop cut Harvest dates can be estimates via satellite • Harvest dates could be used to correct for timing of crop cut Harvest Date 15/5/16 20/5/16 25/5/16 30/5/16 05/6/16 10/6/16
- 11. Problem: Lack of accountability Mobile automated geotagged records can improve accountability and ensure methods • Verify timing • Improve measurement • Verify spatial sampling • Confirm interaction with farmers • ?provide automated feedback to farmers?
- 12. Vegetation Indices NDVI and EVI • “Greenness Indexes” • Vegetation indices are used to monitor vegetation conditions, land cover, land cover changes, and primary production. These data may be used as input to model global biogeochemical and hydrologic processes and global and regional climate.
- 13. Vegetation Indices • Responsive to amount of chlorophyll, leaf area, canopy structure • Healthy or stress plants can be easily identified via satellite or drone
- 14. Problem: Plant health after cut Plant health can be monitored via vegetation indices, or through weather • Adjustments to yield estimates can be made to include disease, water stress etc after crop cuts.
- 15. Problem: Biased location of test plots • Vegetation Indexes can be used to stratify sampling • Strata based on crop stress groups • Area of each strata can be calculated from imagery • Also better accounts for staggered ripening and harvesting Low High Med
- 16. Problem: Translating data to yields/losses Machine Learning With high quality and diverse training data machine learning can integrate data from: • Remote sensing • Ground LIDAR • Weights/measures • Cellphone • Traditional crop cut • Questionnaires • Enumerator quality Yield/Loss Estimate
- 17. Issue: Challenge of Training Data • Ground Truth Data Essential and Largely Missing for Public • Local, contextual ground truth data is going to be required • What is planted, where and when? • Management practices • Plot level yields, crop cuts • Pests, disease • Farmer impressions of loss
- 18. Other Data of Interest In-the-field capture of tenure rights by communities and individuals using mobile devices. Existing tenure data, aerial and satellite imagery can be cached on the device to support data capture in areas with no internet connectivity.
- 19. Other Data of Interest • Open source version of google maps • Anyone can add/edit the global base map • Map plots, farms etc • Getting major support as global base map for ‘unnamed’ competitors of google. Open Street Map
- 20. Other Data of Interest • Allows for field data capture without internet • Basemaps are printed, edited, and scanned back to openstreetmaps.org Walking Papers
- 21. Issue: Alternative yield estimates Research Question • To what degree can we accurately estimate wheat yields for a location over time? Broader Project Objectives 1. Estimate wheat yields at a variety of temporal and spatial scales 2. Develop scalable algorithms, with an eye towards using high resolution imagery in the future Mann, M. L., & Warner, J. M. (2017). Ethiopian wheat yield and yield gap estimation: A spatially explicit small area integrated data approach. Field Crops Research, 201, 60-74.
- 22. Summary Statistics – Compressing Time Properties of a growing season can be summarized in a variety of ways Greener Wheat Rice Wheat Rice Wheat
- 23. Summary Statistics – Maximums, Means etc Values change each season reflecting growing conditions Greener Wheat Rice Wheat Rice Wheat
- 24. Greener Summary Statistics – Area Under the Curve (AUC) Persistence and intensity of greeness Wheat Rice Wheat Rice Wheat
- 25. Greener Summary Statistics – Comparisons to Quantiles How does this year compare to the best years? Wheat Rice Wheat Rice Wheat
- 26. Visualizing Model Performance Within R-Squared: 0.67 Predicted Actual The Take Away Aggregated across districts NDVI by itself can reasonably predict wheat yields over time. Can these tools be applied at the plot level?
- 30. • Plant density • plant spacing • Evenness of the field X • Number of heads • Area measurements • Head length • # grains per head • Grain weight / size • Crop maturity • Assess soil moisture immature crops • Satellite can calculate days until harvested
- 31. After hand threshing, cellphone cameras with machine learning can: • Count kernels • Flag potential disease
- 32. 1 m horizontal resolution 0.1m vertical
- 33. • A developmental main stage when yellow anthers are clearly visible on spikes. It is also called ‘flowering’. Each floret’s lemma and palea are forced apart by swelling of their lodicules, which allows the anthers to protrude. After a day or two, the lodicules collapse and the florets close again. In some circumstances, florets may never show the anthers. When anthers are sterile, as may occur in low- boron soils, the florets may stay open for days, or until cross-pollination occurs. •
- 49. The Possibility of Training Data New low cost tools can be used to collect data on: • Plant health • Cell phone imagery • Plot data through mini surveys • Planting/Harvest dates • Input use • Weather • Disease/pests • Farmer directly linked • Impressions of loss • Mechanism for making a claim As far as I am concerned, this changes everything.
- 50. Future Directions for Remotely Sensed Data Machine Learning & Computer Vision Data from satellites, cellphones, stationary cameras, networked sensors. Monitor: • Yields • Plant growth, height • Pest / Disease • Irrigation systems • Weed management • Row spacing
- 51. Vegetation Indices NDVI and EVI • Responsive to amount of chlorophyll, leaf area, canopy structure • Vegetation indices are used to monitor vegetation conditions, land cover, land cover changes, and primary production. These data may be used as input to model global biogeochemical and hydrologic processes and global and regional climate.