Plant Root Analyzer System

Project Overview

Complete computer vision system developed for the NPEC research facility to automatically monitor plant root growth. The system uses a U-Net deep learning model for root segmentation, individual root tracking over time, and integration with a robot for automated feeding. The full production-ready system includes a CLI tool, FastAPI backend, React dashboard, and Azure cloud deployment with 24/7 availability. The project was executed in two phases: Block B (CV model development and robotics) and Block D (full-stack system development in a team of five).

Challenges

The system needed to detect roots with millimeter precision for robotics applications. Separating and tracking individual roots when they overlap or are complexly intertwined was a major challenge. Additionally, the complete system engineering, including API, monitoring, CI/CD, and cloud deployment, required the same attention as model accuracy. Close collaboration with NPEC biologists to understand the significance of accurate biological measurements was essential.

Results

The U-Net model achieved an 11% SMAPE score for root length predictions, which is exceptionally accurate for biological measurements. The system processes five plants per image with individual tracking. Three deployment options are available: Azure cloud (24/7 multi-location access), on-premise GPU for fast batch processing, and local development.

11%

SMAPE Score (accuracy)

5

Plants per image

24/7

Cloud Availability

100%

Automated Pipeline

Complete Solution

🚀 From Computer Vision to Production

This project demonstrates the complete workflow of an ML system: from data labeling and model training to cloud deployment and monitoring. Not a "proof of concept" but a working system that NPEC researchers could use.

System Features

🔬 Accurate Detection

U-Net architecture for root detection with 11% SMAPE score

📊 Per-Root Tracking

Track up to 5 plants simultaneously with measurements and growth rate per individual root

🤖 Robot Integration

Exact root-tip coordinates for automated feeding system

📈 Growth Analysis

Visualizations of root growth over time with trend analysis per plant

☁️ Cloud Deployment

Azure Container Apps with auto-scaling and 24/7 availability

🔄 Auto-Training

Automatic model updates with new data via feedback loop

U-Net segmentation with root-tips (blue)

Individual root growth tracking dashboard

Technical Implementation

Technology Stack

🧠 U-Net Model

🔥 TensorFlow

⚡ FastAPI

⚛️ React

🐳 Docker

☁️ Azure Cloud

🔄 GitHub Actions

📦 Poetry

System Architecture

The system consists of multiple components that together form a complete solution:

CLI Tool: Command-line tool for batch processing of plant images
REST API: FastAPI backend for image upload, predictions, and results
Web Dashboard: React frontend with real-time visualizations and export functionality
Training Pipeline: Automatic retraining with new data and model updates
Cloud Infrastructure: Azure Container Apps with auto-scaling and monitoring
On-premise Deployment: Docker containers for local inference on NPEC hardware

Deep Learning Pipeline

Model: U-Net with 256x256px input, 4 encoder/decoder layers, and skip connections for detail retention
Training Data: Manually annotated datasets by students, including detailed labeling of roots, leaves, and seeds for accurate model training
Performance: 11% SMAPE (Symmetric Mean Absolute Percentage Error) for root length
Post-processing: Connected components for root identification and skeleton extraction for tip detection
Output: Segmentation mask, root-tip coordinates (x,y pixels), length measurements (mm), and confidence scores

🎯 Why U-Net?

U-Net was specifically developed for biomedical image segmentation and performs well with limited training data. The skip connections ensure the model retains both fine details (precise edges) and global context (correct classification). This makes U-Net perfectly suited for root detection, where both accurate edges and overall root structure are crucial.

Deployment & Infrastructure

☁️ Cloud API

Azure Container Apps
Auto-scaling enabled
99.9% uptime SLA

🖥️ Web Interface

React Dashboard
Real-time updates
Interactive charts

🏢 On-Premise

Docker Deployment
GPU support
Local processing

🔄 CI/CD

GitHub Actions
Automated testing
Zero-downtime deploys

Software Engineering

Code Quality: Flake8 linting for code style and pre-commit hooks
Testing: Unit tests with pytest and integration tests for API, >80% code coverage
Dependencies: Poetry for reproducible builds and dependency management
Documentation: Complete API docs via FastAPI/Swagger and inline code comments
Version Control: Git with feature branches, pull request reviews, and semantic versioning
Monitoring: Azure Application Insights for performance and error tracking

Deployment Options

The system supports three deployment scenarios:

Cloud Production: Azure Container Apps for external access, used by multiple NPEC locations
On-premise GPU: Docker containers on local machines with GPU for faster processing of large batches
Local Development: Development environment with hot-reload for fast iteration during development

Results & Impact

11%

SMAPE Score

5

Plants per image

24/7

Availability

100%

Automation

NPEC Feedback & Usage

"NPEC was very satisfied with the system's implementation. They were particularly enthusiastic about the ability to monitor and visualize individual root growth over time. The combination of accurate segmentation and per-root tracking provides valuable insights for their research context. The system demonstrates well how computer vision can be applied within plant phenotyping."

Specific impact:

Time savings: From manual measurements to automated analysis
Consistency: Constant measurements without variation between different observers
Scalability: Possible to track hundreds of plants in parallel
Robotics integration: Accurate coordinates enable automated feeding
Research value: Time-series data from individual roots opens new research possibilities

System Output per Upload

For each uploaded plant image, the system generates:

Segmentation masks: Red overlay showing exact root locations
Tip coordinates: Blue markers on root-tips with (x,y) pixel coordinates for robotics
Length measurements: Calculated root length in millimeters per root via skeleton path length
Growth rate: Growth rate calculations when multiple timepoints are available
CSV export: Structured data export for statistical analysis in R/Python
Confidence scores: Model certainty per root for quality control

Project Evolution: From CV to Complete System

📅 Block B (January - March 2024): Computer Vision Foundation

Development and training of U-Net segmentation model
Setting up data annotation pipeline
Integration with robotic feeding system for tip coordinates
Root-tip detection algorithm with connected components
Model evaluation and hyperparameter tuning

📅 Block D (April - June 2024): Complete System

React dashboard with interactive visualizations
FastAPI backend with RESTful endpoints
Cloud deployment on Azure Container Apps
Automatic training pipeline with feedback loop
CI/CD setup with GitHub Actions for testing and deployment
Production monitoring and error tracking

Team Collaboration

👥 From Solo to Team Project

What started as an individual computer vision project in Block B grew into a full team effort in Block D. With 5 team members, we specialized. My primary focus was on the deep learning pipeline, training automation, and system architecture.

Technical Challenges & Lessons

Data quality is crucial: Time and care during annotations pays off in better model performance. Inconsistent labels reduce model accuracy.
U-Net skip connections are powerful: The combination of detailed edges and context makes U-Net extremely suitable for biological structures with variable shapes.
Production encompasses more than the model: Engineering around the model, such as APIs, monitoring, and CI/CD, is just as important for practical use as the model itself.
Cloud costs require planning: Azure Container Apps with good scaling policies keep costs manageable, while always-on instances can quickly become expensive.
Domain expertise is essential: Collaboration with NPEC biologists was crucial to understand what "accurate" root measurements mean in a research context.
Feedback loops improve models: The system allows researchers to submit corrections, enabling continuous model improvement.

What Would I Do Differently?

Invest more time in the annotation process: better guidelines for consistent labeling could have further improved model performance.
Test different model architectures in parallel: we started early with U-Net, while other architectures like SegNet would have been interesting to compare.
Write documentation during building instead of afterwards.
Track model versions and experiments from day one: with tools like Weights & Biases we would have had better overview of which hyperparameters worked best.

Project Overview

Challenges

Results

Complete Solution

🚀 From Computer Vision to Production

System Features

🔬 Accurate Detection

📊 Per-Root Tracking

🤖 Robot Integration

📈 Growth Analysis

☁️ Cloud Deployment

🔄 Auto-Training

Technical Implementation

Technology Stack

System Architecture

Deep Learning Pipeline

🎯 Why U-Net?

Deployment & Infrastructure

☁️ Cloud API

🖥️ Web Interface

🏢 On-Premise

🔄 CI/CD

Software Engineering

Deployment Options

Results & Impact

NPEC Feedback & Usage

System Output per Upload

Project Evolution: From CV to Complete System

📅 Block B (January - March 2024): Computer Vision Foundation

📅 Block D (April - June 2024): Complete System

Team Collaboration

👥 From Solo to Team Project

Technical Challenges & Lessons

What Would I Do Differently?

Check Out My Other Projects