Context
Abelio, a leader in digital farming solutions, uses computer vision to process aerial images, like drone and satellite images, to provide farmers with actionable insights that help navigate the challenges of modern agriculture.
As their computer vision operations scaled, Abelio faced some bottlenecks:
- Managing increasing data volumes during peak farming seasons.
- Rapidly retraining models to meet tight deadlines (48-hour turnaround from image acquisition to insights).
- Ensuring traceability and reproducibility for experiments across diverse environments.
Abelio overcame these challenges with Picsellia; improving data organization, accelerating retraining processes, and ensuring model stability. Our Computer Vision platform allowed them to expand their impact, enabling more farmers to make informed decisions.
About Abelio
Abelio partners with agriculture cooperatives to leverage the power of aerial imagery and AI, transforming images into actionable insights for farmers. By processing drone and satellite images captured, Abelio supports precision agriculture, helping farmers address challenges like fertilization, irrigation, disease prevention, and weed detection.
For example, Abelio can help farmers apply just the right amount of fertilizer to maximize crop profitability while minimizing negative ecological impact. Abelio can also help optimize water use during critical growth periods, estimate crop health to prevent disease outbreaks and identify weeds efficiently to reduce treatment costs.
Abelios CV Journey
Initially, Abelio relied on AWS tools like S3 for data storage, EC2 for model training, and SageMaker for deployments. While functional, this fragmented setup had several issues:
- AWS S3 lacked tools for image visualization, complicating dataset management.
- Scaling operations during peak seasons created inefficiencies in handling large image volumes.
- Retraining models to meet tight delivery deadlines was time-intensive and prone to errors.
As their activity grew, Abelio felt the need for an integrated MLOps platform for computer vision. During peak seasons, they faced a fourfold increase in image processing, requiring a solution to handle storage, organization, and workflow management effectively.
Key improvements with Picsellia
- Centralized data management Simplified image storage, retrieval, and organization.
- Efficient retraining Reduced retraining cycles to 48 hours.
- The annotation campaign tool Enabled progress tracking, quality control, and independent annotation workflows.
These advancements allowed Abelio to focus on improving model performance, delivering insights faster, and scaling their operations.
Business Impact
Picsellia’s integrated features have transformed Abelio’s approach to computer vision:
- Time-to-model: By streamlining dataset management and retraining workflows, Abelio reduced retraining time to 48 hours, meeting the tight timeframe for farmers.
- Model accuracy: Improved dataset reliability and tracking have stabilized models, increasing average precision and recall.
- Data volume handling: Picsellia has allowed Abelio to manage the difficult nature of data seasonality efficiently.
- Annotation efficiency: The annotation campaign feature has helped Abelio oversee campaign progress, improve annotation quality, and streamline team workflows.
With Picsellia, Abelio has also enhanced model tracking, ensuring reproducibility and traceability by recording dataset versions, base models, and training parameters. This has led to better analysis and comparison of experiments, reducing errors and improving reliability.
Moving Forward
With Picsellia’s scalable and collaborative platform, Abelio has positioned themselves to expand their offerings in precision agriculture. The platform enables R&D and production teams to work together seamlessly, allowing Abelio to respond quickly to their clients' needs.
