Skip to content

My Original Research Contributions

This page documents my original contributions to 3DGS-based plant phenotyping.

1. Scale-Invariant Height Extraction

The Problem

Traditional PLY-based height measurement suffers from temporal scale inconsistency.

PLY Method Issues Figure 1: Scale inconsistency in PLY-based measurement across dates (CV: 28.0%)

Challenge Identified

Structure-from-Motion (SfM) reconstruction produces different coordinate scales for each capture date, making direct height comparison unreliable.

My Solution

Innovation: Rendered image-based trait extraction in normalized image space, used for relative growth monitoring rather than absolute measurement.

graph TD
    A[3DGS Model] --> B[Render Top View]
    B --> C[Normalize by Image Dimensions]
    C --> D[Extract Height in Pixels]
    D --> E[Scale-Invariant Measurement]

    style E fill:#90EE90

Implementation Video

Video 1: Demonstration of scale-invariant height extraction method

Results Comparison

PLY vs Rendered Figure 2: My method achieves 18.2 percentage point (2.86×) improvement (CV: 28.0% → 9.8%)

Original Contribution #1

First demonstration of scale-invariant trait extraction from multi-date 3DGS reconstructions for plant phenotyping.

  • Method: Image-space normalization
  • Result: 18.2pp CV improvement (2.86×)
  • Validation: 22 dates, 49 days
  • Impact: Enables reliable time-series growth monitoring (relative change, not absolute trait values)

Calibration is the wrong fix

Applying a per-session scale factor to the direct-3D height makes it worse, not better (CV 35.1% vs 28.0%): the estimated scale factor is itself noisy and injects its own variance. This is what motivates measuring in the render, where the scale cancels by construction.

Biological Validation — Pruning-Event Detection

The strongest evidence that the render-space signal tracks real biology (not just numerical stability) is that it detects documented crop-management events.

Pruning events aligned with trait drops Render-space height over 49 days: smooth regrowth punctuated by three sharp drops, each aligned with a documented pruning event.

Biological validation

Three documented pruning events are each detected as a sharp drop in render-space height (mean drop ≈ 0.20 in normalized height, 15–27 percentage points — well above the 9.8% background variation).

  • Significance: p = 0.0008 (drops at pruning dates vs. non-pruning dates)
  • Meaning: the pipeline responds to real management events, so it is a usable monitoring / change-detection signal.

Physical-Reference Corroboration — 45 cm Pipe

To ground the render-space height against something physical, an in-scene bench pipe of known length (45 cm) is used as a ruler that is co-visible with the plant.

Reference height: h_gt = (plant_px / pipe_px) × 45 cm, annotated in all 22 sessions.

Physical-reference agreement

Render-space height correlates with the pipe-ratio reference across all 22 sessions:

  • Pearson r = 0.74 (p < 0.001), R² = 0.55
  • Reference mean height 159.9 cm; comparable spread (h_gt CV 8.9% vs. h_norm CV 9.6%)
  • The pipe spans 263–420 px for a fixed 45 cm, so the ~45% unexplained variance is noise in the reference instrument (SfM scale ambiguity), not pipeline error.

Honest scope

This is corroboration that the signal is grounded in reality — not proof of absolute accuracy. No calibrated tape-measure ground truth (RMSE) was collected for these sessions; see Results & Validation for the limitation and the planned ground-truth protocol.

2. Environmental Correlation Analysis

Sensor Deployment

Greenhouse Setup Photo 1: Multi-modal sensor deployment in greenhouse environment

Data Collection

I integrated IoT sensors measuring: - Temperature (°C) - Humidity (%) - Solar radiation (W/m²)

Discovery: Humidity Correlation

Correlation Matrix Figure 3: Environmental factors vs 3DGS reconstruction quality (n=18)

Original Discovery

Significant humidity correlation with PSNR:

  • Correlation coefficient: r = +0.506
  • P-value: p = 0.032 (significant at α=0.05)
  • Interpretation: Higher humidity → better reconstruction quality

First identification of environmental effects on 3DGS quality

Radiation-Based Classification

Radiation Classification Figure 4: Data-driven 100 W/m² threshold with WMO validation

My Method: 1. Data-driven: Natural gap at 57.6 W/m² (47.1 → 104.7) 2. WMO standard: 100 W/m² meteorological threshold 3. Statistical: Perfect balance (n=9 vs n=9)

Original Contribution #2

First radiation-based classification for 3DGS reconstruction quality in controlled environments.

3. Complete 50-Day Validation

Time-Series Dataset

PSNR Over Time Figure 5: Temporal stability across 22 dates (PSNR: 23.84 ± 0.83 dB, CV: 3.5%)

Growth Monitoring Results

Growth Curve Figure 6: 49-day continuous monitoring with my pipeline

Time-lapse Video

Video 2: 49 days of tomato growth captured with my 3DGS pipeline (Jan 19 - Mar 9, 2026)

Original Contribution #3

First long-term validation of 3DGS for time-series plant phenotyping.

  • Duration: 49 days
  • Frequency: 22 capture dates
  • Consistency: CV = 3.5% (PSNR)
  • Growth tracking: Positive correlation (r = 0.209)

4. Complete Pipeline Integration

System Architecture

Pipeline Architecture Figure 7: Complete system architecture integrating 3DGS with IoT sensors

Execution Demonstration

Video 3: Complete pipeline execution (small video hosted directly)

📊 Summary of Contributions

Contribution Innovation Impact Validation
Scale-invariant render-space monitoring Image-space normalization 18.2pp CV improvement (2.86×) 22 dates, 49 days
Pruning-event detection Change detection from render-space traits Biological validation 3 events, p = 0.0008
Physical-reference corroboration In-scene 45 cm pipe ratio Grounded in real-world scale r = 0.74, p < 0.001
Environmental Correlation Humidity-PSNR relationship First identification r=+0.506*, p=0.032
Radiation Classification Data-driven 100 W/m² threshold WMO-validated method Perfect balance n=9:9
Long-term Validation 49-day continuous monitoring Temporal consistency CV = 3.5%

📄 Publications

My research is documented in:

✅ Progress Report 4 (April 2026)
Complete methodology and validation results

Download PDF

📝 Master's Thesis (September 2026)
In progress

Target submission: September 2026

📄 Computers and Electronics in Agriculture (Elsevier)
Under review

Scale-invariant 3DGS for time-series greenhouse plant monitoring

🎓 Presentations

  • ✅ ISFAR-SU 2026 (Completed) - Domestic conference presentation
  • 🎯 International Conference (Planned) - Target: Late 2026

All figures, videos, and results shown on this page are original work conducted by Zobaer Al at Mineno Laboratory, Shizuoka University.