Harvest Estimation Tool:
Objective:
Manual counting in a subregion of the entire farm and subsequent extrapolation is the general practice in blueberry farms to estimate berry counts. This is not practical and less accurate when dealing with large scale blueberry farms. Poor accuracy in estimation can lead to poor planning for selling and wastage of berries. In this project, I led a team of 3 engineers to develop a Machine Learning based tool for estimating location and counts of 7 different stages of blueberries across the farms and subsequent harvest estimations.
Contributions:
- Trained a multi-head Unet algorithm for segmentation and count estimation of 7 stages of blueberries using a custom objective function focusing on handling occlusion and overlapping among berries.
- Contributed to training an LSTM model to forecast berry counts for upcoming dates using historical data that includes demographic and environmental conditions from previous days.
- Led the development of a cloud based single client application for generating report with harvest estimation and berry heatmap.
Result:
- Successfully deployed the cloud based application on the client's field for harvest estimation.
- Results from initial testing indicated a 50% improvement in berry count estiamtion compared to the traditional method.
- Availability of heatmap in addition to berry counts allowed the client to plan berry picking route improving overall efficiency.