GeoAwesome Meetup is the web’s premier digital networking event for the geospatial community. At this year’s event, Dimitris Zermas, Principal Scientist at Sentera and iMerit’s VP of Strategic Business Development Jai Natarajan spoke on the deployment of machine learning for geospatial analysis of agriculture data.
Here are 5 key takeaways:
- Data acquisition and data curation present the main challenge in the deployment of scalable machine learning in agriculture, accounting for 80% of the problems. The selection of appropriate cloud computing tools takes up the remaining 20%.
- Edge cases can make or break the case for AI adoption in precision agriculture. They are the last mile problem, responsible for reducing the accuracy of algorithms by up to 20%. When Sentera faced similar issues, iMerit’s fusion approach helped improve algorithm performance by 19%, while maintaining ongoing annotation scaling and growth. Pay more attention to the last mile edge cases to improve the chances of AI adoption.
- Human expertise is indispensable in training AI for precision agriculture, especially for biological observations. Data selection, annotation, and quality control all require individuals trained/specializing in agronomic technologies.
- Success in the labs does not always translate to success in the field. Algorithms often deliver less than 100% accurate results in field deployments. In business-critical processes, even a deviation by 5% can have major economic ramifications.
- Sensors, drones, and other sources deliver a constant stream of real-time data. Without high-quality data audits and frequent refreshing of input data, the algorithm may fail to spot relevant changes like an increase in pests/weeds, or even misdiagnose healthy crops as diseased.