This repository contains the data and code used in the study:

“Leveraging Foundation Models and Time-Lapse Photography for Semi-Automated Tracking of Tree Apical Growth”

---

🌲 Project Overview

Primary height growth determines tree seedling survival, competition, forest structure, and productivity. However, efficient in-situ methods to track fine-scale height growth over time are lacking.

This project introduces a semi-automated framework that combines:

• Time-lapse photography
• The Segment Anything Model v2 (SAM2)
• Skeleton-based length measurement

to quantify intra-seasonal conifer apical growth.

The method was applied to 105 balsam fir leaders from two sites in Québec, Canada, spanning 13 growing seasons.

---

📁 Repository Structure

Data.zip/

├── AnnualROICroppedDailyImages/ # Cropped time-series images

├── VideoSAM/ # empty folder for SAM2 outputs (frames, prompts, masks, results)

├── Measurements/ # Manual reference measurements in XLSX and CSV format (code uses XLSX files)

├── AblationSensAnalysis/ # Results of ablation sensitivity analysis (9792 configurations)

Code.zip/

└── LengthTracker.py # Main processing script

---

⚙️ What the Code Does

LengthTracker.py performs the full workflow:

1. Manual input
   • Draw bounding boxes around leaders (every 5 frames)
   • Select the transition between early (bud/whorl) and late (leader extension) growth phases
2. Automated processing
   • Propagate segmentations through time using SAM2
   • Clean masks (noise removal, region filtering)
   • Measure leader length using skeleton-based curved path extraction
   • Detect and interpolate outliers
   • Apply seven-day rolling median window smoothing
3. Output
   • Daily leader length (pixels)
   • Percent of seasonal maximum growth
   • Comparison with manual measurements stored in annual XLSX files in "Measurements" folder
   • CSV files for each leader

---

💻 System Requirements

This workflow requires a CUDA-enabled GPU.

Tested with:

• Python 3.10+
• PyTorch with CUDA
• SAM2 (Meta AI Research)

---

📦 Dependencies

Install SAM2 following the official instructions:
https://github.com/facebookresearch/sam2

Additional Python packages used:

numpy
pandas
opencv-python
torch
torchvision
matplotlib
scikit-image
networkx
scipy
statsmodels
openpyxl
Pillow

---

🚀 How to Run

Step 1 — Install SAM2

Follow the official installation guide from Meta AI Research (https://github.com/facebookresearch/sam2).

Step 2 — Set Model Paths

In LengthTracker.py, update:

sam2_checkpoint = "/path/to/sam2.1_hiera_large.pt"
model_cfg = "/path/to/sam2.1_hiera_l.yaml"

---

Step 3 — Organize Data

Ensure the directory structure matches:

Data/AnnualROICroppedDailyImages/{site}/{leader_ID}/

---

Step 4 — Run the Script

python Code/LengthTracker.py

---

✋ Manual Steps Required

The workflow is semi-automated. Users must:

1. Draw a box around the leader every 5 frames
2. Select the start of Phase 2 (apical dominance)

These prompts are saved and then used for segmentation.

---

📤 Outputs

For each leader:

VideoSAM/{site}/{leader_ID}/Results/
└── leader_{ID}_manual_vs_sam_lengths.csv

Contains:

• Manual leader length (px)
• SAM vertical length (px)
• SAM curved length (px)
• Percent seasonal growth

---

📊 Validation Summary

• SAM2 lengths closely matched manual measurements (R² = 0.98)
• Semi-automated workflow was ~20× faster than manual measurement

---

📜 Citation

If you use this code or data, please cite:

Spafford et al., Leveraging Foundation Models and Time-Lapse Photography for Semi-Automated Tracking of Tree Apical Growth, [Ecological Informatics, 2026]

---