Sebastian Kleiner
Data Science Fellows June 2020 Cohort
Abstract
CropX models constantly adapt to the crop’s growth stage and to the changing conditions of the soil and weather, to provide crop-specific recommendations that help achieving maximal yield, validated in the field to generate a crop-yield enhancement.
To achieve this solution, CropX needs to constantly validate its data in order to give accurate results and insights to its users. This is why the Data Science team decided to create a model for field identification based on its NDVI behavior. This metric uses the Vegetation Index across time and it’s extracted from satellite images. By analyzing the pattern of each field, a clustering model is created to distinguish the obvious-routine ones from the noisy fields (planting potatoes every year or switching crops). This will allow CropX to validate past data and understand better the field’s soil.
Afterwards, an outlier detection model is constructed on top of the obvious-routine fields to classify each season or cycle inside fields.
Challenges (at least two)
- Rapidly understand CropX culture, systems and databases while achieving results independently.
- WFO in an independent environment.
Achievements (according to KPIs)
- Clustering 80% of fields that have been eye-labeled as ‘Obvious-pattern’ correctly.
- 67% of outlier counting detection correctly assigned with an unsupervised model contrasted with eye-labeled data.
Future project development
- Improve KPIs using alternative features from the times series.
- Establish an end-to-end process towards the production environment.
