Microplastic Detection & Geospatial Mapping
Leveraging the discovery that moss can indicate microplastic presence, this challenge focuses on developing a robust computer vision and data science pipeline. Participants will build an end-to-end system capable of identifying and quantifying microplastic particles from microscopic imagery of moss samples. The solution should integrate a machine learning model for object detection (e.g., YOLO or Mask R-CNN), process image data to extract quantitative metrics (size, count), and then utilize geospatial data to map contamination levels. This project simulates a real-world environmental monitoring system, requiring careful data management, model training, and results visualization. The technical implementation will involve modern MLOps practices, orchestrating the data ingestion, preprocessing, model training, and inference steps using a workflow management system like Dagster. Further, the solution will incorporate the Llama 3.2 model to assist in contextualizing results, such as generating descriptive summaries of contamination findings for specific geographic areas, or suggesting potential sources based on observed microplastic types. This challenge emphasizes not only accurate detection but also scalable deployment and actionable reporting for environmental scientists and policymakers.
What you are building
The core problem, expected build, and operating context for this challenge.
Build an ML pipeline using computer vision to detect microplastics in moss imagery, map them geospatially, and generate interpretive summaries with Llama 3.2, orchestrated by Dagster.
How work is evaluated
The solution will be evaluated on the accuracy of microplastic detection, the robustness and reproducibility of the Dagster pipeline, the quality of geospatial mapping, and the coherence and insightfulness of Llama 3.2 generated reports.
Challenge datasets
Public datasets, plus the private datasets you have access to.
What you should walk away with
Learning objectives will be added soon
Use the overview and evaluation guide as the source of truth for expected outcomes.
Reference links and supporting material
Sample data for 3 tasks
Challenge setup
The solution will be evaluated on the accuracy of microplastic detection, the robustness and reproducibility of the Dagster pipeline, the quality of geospatial mapping, and the coherence and insightfulness of Llama 3....
Challenge input
A directory containing high-resolution microscopic images (PNG/JPG) of moss samples, potentially with ground truth bounding boxes/masks in COCO for...
Agent execution
No agent tool is listed for this challenge.
Evaluated output
A JSON file containing detected objects with bounding boxes, segmentation masks, confidence scores, and predicted class for each image. Additionall...
- Verifies that the Dagster pipeline executes successfully without errors and generates all expected intermed...
- Confirms that the generated GeoJSON is valid and contains features with appropriate geometries and properties.
- Ensures the generated report by Llama 3.2 meets a minimum length requirement to be comprehensive.
- MAP (Mean Average Precision) target: 0.75
- Python execution harness
Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.
View technical recipe
Configured tools
No tool records are attached.
Evaluation contract
- Verifies that the Dagster pipeline executes successfully without errors and generates all expected intermediate and final outputs.
- Confirms that the generated GeoJSON is valid and contains features with appropriate geometries and properties.
- Ensures the generated report by Llama 3.2 meets a minimum length requirement to be comprehensive.
Metrics
- mAP (mean Average Precision) · Target 0.75
- Average Microplastic Count Error · Target 0.1
- Geospatial Hotspot Accuracy · Target 0.6
- LLM Report Factual Consistency (RAGAS-like) · Target 0.8
Workflow tasks
- Task 1
Input: A directory containing high-resolution microscopic images (PNG/JPG) of moss samples, potentially with ground truth bounding boxes/masks in COCO format.
Output: A JSON file containing detected objects with bounding boxes, segmentation masks, confidence scores, and predicted class for each image. Additionally, a CSV with image_id, microplastic_count, average_size_pixels.
Execution harness: Python
- Task 2
Input: A CSV file with image_id, latitude, longitude, and the output JSON/CSV from 'MicroplasticDetection'.
Output: An HTML file containing an interactive map (e.g., Folium) showing microplastic hotspots and a GeoJSON file of aggregated contamination data.
Execution harness: Python
- Task 3
Input: The aggregated microplastic metrics (CSV), geospatial analysis results (GeoJSON), and a set of predefined prompts for Llama 3.2.
Output: A Markdown or PDF document containing the generated natural language report.
Execution harness: Python
Recipe freeze marker
No recipe freeze time is recorded.