The xDF-GOG project provides advisory solutions for farmers worldwide, helping them manage weeds and optimize crop protection using advanced drone technology. The entire solution is developed on AWS, leveraging cutting-edge services such as Lambda Functions, Step Functions, Glue, RDS, S3, and AWS Athena to process and query vast drone imagery datasets efficiently.

As Machine Learning Engineers, our team faced the complexity of managing a distributed pipeline for processing drone-captured agricultural imagery. The workflow required seamless coordination of batch processing jobs, serverless Lambda functions, and dynamic decision-making logic.



Key hurdles included ensuring fault tolerance across heterogeneous components, scaling to handle large geospatial datasets, and integrating outputs with an existing Field Management backend service via RabbitMQ. Farmers needed real-time, actionable insights (e.g., weed heatmaps, spray zones) to optimize crop management, but manual orchestration introduced latency and reliability risks.

We implemented a serverless orchestration framework using AWS Step Functions to automate the end-to-end workflow. The state machine architecture included:

• Image ingestion: Triggered by S3 uploads of drone imagery from farmers.
• Parallel processing: Lambda functions for image normalization, metadata extraction, and quality checks.
• Batch inference: Scalable containerized jobs running ML models for weed detection and spray recommendations.
• Decision nodes: Conditional routing based on confidence thresholds or regional farming policies.
• Output integration: Final maps (geoTIFF/GeoJSON) published to the backend via RabbitMQ using a custom Lambda-SQS bridge.

Step Functions’ built-in retries, error handling, and state tracking ensured resilience, while its visual interface simplified monitoring. By decoupling pipeline stages and leveraging AWS service integrations (S3, Lambda, Batch), we achieved 40% faster processing cycles and 99.8% system uptime. Farmers now receive precision agriculture maps within hours, enabling targeted herbicide use and reducing environmental impact by approximately 15%.