Sheep Dietary Analysis ML System

Overview

An ML pipeline developed in collaboration with Texas A&M AgriLife Research (Dr. Walker's lab) that automates the analysis of microhistological samples to determine sheep dietary composition. The system processes microscope imagery to identify plant species fragments in fecal samples.

Problem

Understanding what livestock eat on rangelands is crucial for:

• Optimizing grazing management strategies
• Monitoring rangeland health over time
• Studying animal nutrition and forage preferences

Traditional microhistological analysis is labor-intensive: trained technicians manually examine hundreds of plant fragments under a microscope and identify species based on cellular structures. This process is:

• Time-consuming (hours per sample)
• Subject to inter-observer variability
• Difficult to scale for large research studies

Approach

Data Pipeline

Built an end-to-end pipeline that:

1. Image Preprocessing: Standardizes microscope images (color normalization, noise reduction, segmentation)
2. Feature Extraction: Extracts relevant visual features from plant fragment images
3. Classification: Identifies plant species using trained ML models
4. Aggregation: Computes dietary composition percentages across sample sets

Model Development

Experimented with multiple approaches:

• Traditional ML with handcrafted features (texture, shape, color histograms)
• Transfer learning from pretrained image models
• Ensemble methods combining multiple feature types

The final system uses an ensemble approach that balances accuracy with interpretability—important for research applications where understanding model decisions matters.

Validation Strategy

• Cross-validation against expert-labeled reference samples
• Comparison with traditional manual analysis results
• Sensitivity analysis across different image quality conditions

Stack

| Component | Technology | |-----------|------------| | Language | Python | | ML | scikit-learn, XGBoost | | Image Processing | OpenCV, scikit-image | | Data | Pandas, NumPy | | Visualization | Matplotlib, seaborn |

Results

• Automated species identification for common rangeland plant fragments
• Reduced analysis time compared to fully manual methods
• Provided consistent, reproducible classifications across samples

Key Learnings

1. Domain expertise is essential — Close collaboration with researchers ensured the model solved the right problem and outputs were interpretable
2. Data quality trumps model complexity — Clean, well-labeled training data mattered more than sophisticated architectures
3. Research applications need explainability — Black-box predictions weren't enough; researchers needed to understand and validate model decisions

Context

This project is part of ongoing research at Texas A&M AgriLife Research under Dr. Walker's supervision. The system supports studies on livestock grazing behavior and rangeland ecology.