Ari Salkow
Data Science Fellows February 2021 Cohort
Abstract
The objective of the project was to build a machine learning classifier that could tell whether a specific piece of land in Mali had been eaten by a pest called The Fall Armyworm. The base model would be binary i.e. had been eaten or not. As a stretch target, the model should predict whether a piece of land had been eaten, and the severity of the damage. The data to be used was:
- Appropriate freely available satellite imagery – a field called Remote Sensing.
- Ground-truth data from the FAO – an organization that provides such information.
Challenges (at least two)
- To execute this. It was necessary to apply CV techniques that were a modified version of the ones we learnt at ITC. Through reading papers, watching videos, reading online and learning from co-workers, we greatly improved our understanding and ability for this sub-section of machine learning.
- There were quite a few libraries that we hadn’t encountered. Using the same ways as point 1., we learnt more about these libraries and approaches.
Achievements (according to KPIs)
- We understood this specific insurance industry and Remote Sensing much better by the end of the project.
- We built a segmentation model, albeit using different data, for multi-class image segmentation. This same approach could be applied with the libraries discussed above to build a model for this project, using the requisite data sources.
- Improved Python, data analysis and data science skills by working on the project and collaborating with the other team members from OKO and Omdena.
Future project development
- Various other models and modelling approaches should be tried to improve prediction accuracy.
- Other data suppliers should be explored, to improve performance.
- Perhaps additional data from existing OKO customers could be extracted.
- More research papers, and similar, could be examined to get ideas for improvement.