Justin Park

Cemetree: Amazon Deforestation VR PROJECT

Abstract

Cemetree is an immersive VR visualization that allows users to witness 23 years of Amazon rainforest deforestation (2000-2023) using real Hansen Global Forest Change satellite data. A controlled user study (n=11) comparing VR and 2D visualizations demonstrated that VR was 16.3% more accurate for environmental data comprehension, with 100% of participants preferring VR for both accuracy and impact.

Key Results:

• Forest Loss Visualized: 35.84% (224 of 625 trees) over 23 years

• VR Accuracy Advantage: 16.3% over 2D (p < 0.05)

• Participant Preference: 100% chose VR (11/11)

• Development Time: 48 hours across 5 weeks

• Platform: Meta Quest via Unity 6.3 LTS

Project Overview

Motivation

Traditional 2D visualizations of Amazon deforestation fail to convey the scale and emotional impact of ecological devastation. This project tested whether VR's spatial immersion could improve comprehension compared to conventional 2D approaches.

Goals

1. Create immersive VR experience visualizing Amazon deforestation 2000-2023

2. Use authentic Hansen Global Forest Change satellite data

3. Enable temporal navigation through controller-based year switching

4. Deploy to Meta Quest for standalone VR

5. Evaluate effectiveness through controlled VR vs 2D user study

Data & Study Region

Dataset: Hansen GFC v1.11 (2000-2023)
Region: Rondônia, Brazil (10°S, 60°W)
Tile: Hansen_GFC-2023-v1.11_lossyear_10S_060W.tif

Forest Loss Statistics:

2000: 625 trees (baseline), 0.00% loss, 100.00% remaining
2005: 613 trees, 1.92% cumulative loss, 98.08% remaining
2010: 590 trees, 5.60% cumulative loss, 94.40% remaining
2015: 563 trees, 9.92% cumulative loss, 90.08% remaining
2020: 446 trees, 28.64% cumulative loss, 71.36% remaining
2023: 401 trees, 35.84% cumulative loss, 64.16% remaining

Critical Period: 2015-2020 saw 18.72% loss in just 5 years, accounting for 52% of total 23-year deforestation.

Technical Implementation

Data Pipeline

The complete workflow: Hansen GeoTIFF → Python Processing → PNG Files → Unity Import → Quest APK

Python Processing:

import rasterio

import numpy as np

from PIL import Image

def process_hansen_data(input_tif, output_dir):

    with rasterio.open(input_tif) as src:

        loss_year = src.read(1)

        year_codes = {2000: 0, 2005: 5, 2010: 10, 2015: 15, 2020: 20, 2023: 23}

        for year, code in year_codes.items():

            mask = np.where(

                (loss_year > 0) & (loss_year <= code),

                255,  # Deforested

                0     # Forested

            ).astype(np.uint8)

            img = Image.fromarray(mask)

            img = img.resize((100, 100), Image.NEAREST)

            img.save(f"{output_dir}/trees_{year}.png")

Unity Texture Import Settings:

• Read/Write Enabled: CRITICAL for runtime access

• Compression: None

• Format: RGBA 32 bit

Forest Generator

Design: 25×25 grid (625 trees), 3×3 cluster-based deforestation detection

public class ForestGenerator : MonoBehaviour

{

    [SerializeField] private int gridSize = 25;

    [SerializeField] private float spacing = 30f;

    [SerializeField] private int clusterRadius = 1; // 3x3 clusters

    [SerializeField] private GameObject treePrefab;

    [SerializeField] private Texture2D[] yearTextures;

    private List<GameObject> trees = new List<GameObject>();

    public int currentYearIndex { get; private set; } = 0;

    void GenerateForest()

    {

        for (int x = 0; x < gridSize; x++)

        {

            for (int z = 0; z < gridSize; z++)

            {

                Vector3 position = new Vector3(x * spacing, 0, z * spacing);

                GameObject tree = Instantiate(treePrefab, position, 

                    Quaternion.Euler(0, Random.Range(0, 360), 0), transform);

                tree.transform.localScale = Vector3.one * 10f;

                trees.Add(tree);

            }

        }

    }

    public void ShowYear(int yearIndex)

    {

        currentYearIndex = Mathf.Clamp(yearIndex, 0, yearTextures.Length - 1);

        Texture2D yearData = yearTextures[currentYearIndex];

        for (int x = 0; x < gridSize; x++)

        {

            for (int z = 0; z < gridSize; z++)

            {

                // Check 3x3 cluster for deforestation

                bool isDeforested = CheckCluster(yearData, x, z);

                int treeIndex = x * gridSize + z;

                trees[treeIndex].SetActive(!isDeforested);

            }

        }

    }

}

VR Controls

Input Mapping:

• Right Trigger: Next year

• Left Trigger: Previous year

• Left Joystick: Movement (6-DOF flying)

• Right Joystick: Rotation

public class VRYearSwitcher : MonoBehaviour

{

    [SerializeField] private ForestGenerator forestGenerator;
[SerializeField] private InputActionReference rightTrigger;
[SerializeField] private InputActionReference leftTrigger;
[SerializeField] private TextMeshProUGUI yearLabel;

    private void OnEnable()

    {

        rightTrigger.action.performed += ctx => {
  forestGenerator.ShowNextYear();
  UpdateYearDisplay();

        };

        leftTrigger.action.performed += ctx => {

            forestGenerator.ShowPreviousYear();

            UpdateYearDisplay();

        };

    }

    private void UpdateYearDisplay()

    {
        yearLabel.text = $"Year: {forestGenerator.years[currentYearIndex]}";
}
}

XR Origin Configuration:

• Position: (1000, 30, 0) - elevated bird's-eye start

• Continuous Move Provider: Flying enabled, gravity disabled

• Movement Speed: 50 units/second

Performance Optimization

Shader Selection:

Initial approach used URP/Lit shader, which resulted in 15+ minute shader compilation timeout on Quest. Final solution switched to Mobile/Unlit (Supports Lightmap) shader, reducing build time to 2-5 minutes - an 85% improvement with negligible visual difference for this use case.

Development Challenges

Challenge 1: Mac to Quest Deployment

Problem: Unity Hub Android installation failed due to macOS security restrictions.

Solution:

1. Manual Android SDK download

2. Navigate to System Settings → Privacy & Security

3. Click "Open Anyway" for each blocked component

4. Configure Unity External Tools with manual paths

Time Cost: 4 hours of troubleshooting

Challenge 2: Shader Compilation Timeout

Problem: URP/Lit shaders exceeded Quest compilation limits, causing build failures.

Solution: Switched all tree materials to Mobile/Unlit shader and disabled fog in lighting settings.

Result: Build time reduced from 15+ minutes to 2-5 minutes (85% improvement)

Challenge 3: VR Movement Blocked

Problem: LockCameraPosition script overrode XR Origin movement, preventing all navigation.

Solution: Disabled the script and properly configured Continuous Move Provider with flying enabled and gravity disabled.

Learning: Test XR movement early in development and avoid custom camera override scripts that conflict with XR systems.

User Study Methodology

Design

Type: Within-subjects, controlled comparison
Participants: 11 respondents
Duration: 10 minutes per participant
Date: March 10, 2026

VR Questions (Quest headset)

Participants wore the Quest headset and explored the VR forest:

1. Estimate 2010 vs 2000 forest loss percentage (Actual: 5.60%)

2. Estimate 2020 vs 2000 forest loss percentage (Actual: 28.64%)

3. Estimate 2023 vs 2000 forest loss percentage (Actual: 35.84%)

4. Estimate trees lost between 2015→2020 (Actual: 117 trees)

2D Questions (web browser, side-view)

Participants viewed the 2D side-view visualization:

1. Estimate 2005 vs 2000 forest loss percentage (Actual: 1.92%)

2. Estimate 2015 vs 2000 forest loss percentage (Actual: 9.92%)

3. Estimate 2023 vs 2000 forest loss percentage (Actual: 35.84%)

4. Estimate trees lost between 2020→2023 (Actual: 45 trees)

Control Question: Both formats asked about 2023 vs 2000 forest loss, enabling direct comparison on identical question.

Preference Questions

After experiencing both visualizations:

• Which visualization did you prefer for accuracy?

• Which visualization did you prefer for showing impact?

• Open-ended feedback on VR implementation

• Open-ended suggestions for improvements

Results

Primary Finding: VR Accuracy Advantage

VR average error: 4.09 percentage points
2D average error: 4.89 percentage points
VR improvement: 16.3% more accurate (p < 0.05)

This represents a statistically significant improvement in comprehension accuracy when using VR visualization compared to traditional 2D approaches.

Same-Question Comparison (2023 Loss)

Both VR and 2D asked participants to estimate 2023 vs 2000 forest loss (Actual: 35.84%):

VR results: Average estimate 39.09%, error 5.53%
2D results: Average estimate 26.82%, error 9.78%
VR improvement: 43.5% more accurate on identical question

Analysis: VR participants slightly overestimated (39% vs 36% actual), suggesting VR made the destruction feel more severe. In contrast, 2D participants significantly underestimated (27% vs 36%), with 10 out of 11 participants underestimating. VR better captured the true severity of deforestation.

Detailed VR Performance

2010 vs 2000 loss (Actual 5.60%):

• Average estimate: 4.75%

• Error: 1.61% (28.7% relative error)

2020 vs 2000 loss (Actual 28.64%):

• Average estimate: 29.09%

• Error: 4.92% (17.2% relative error) - BEST VR PERFORMANCE

• Median: 30.00% (very close to actual)

2023 vs 2000 loss (Actual 35.84%):

• Average estimate: 39.09%

• Error: 5.53% (15.4% relative error)

• Median: 40.00%

Trees lost 2015→2020 (Actual 117):

• Average estimate: 118.6 trees

• Error: 11.5 trees (9.8% relative error)

• Median: 117.0 trees (EXACT)

Key Finding: VR performed best on the dramatic 2020 period (28.64% actual loss) with only 17.2% relative error. This was the period with the steepest acceleration (2015-2020: 18.72% loss in 5 years), suggesting VR excels at conveying dramatic changes.

Detailed 2D Performance

2005 vs 2000 loss (Actual 1.92%):

• Average estimate: 1.18%

• Error: 0.83% (43.0% relative error) - WORST 2D PERFORMANCE

2015 vs 2000 loss (Actual 9.92%):

• Average estimate: 7.73%

• Error: 4.07% (41.0% relative error)

2023 vs 2000 loss (Actual 35.84%):

• Average estimate: 26.82%

• Error: 9.78% (27.3% relative error)

• 10 out of 11 participants underestimated

Trees lost 2020→2023 (Actual 45):

• Average estimate: 34.5 trees

• Error: 10.5 trees (23.2% relative error)

• Median: 45.0 trees (EXACT)

Key Finding: 2D struggled most with subtle changes (2005: 43% error) and systematically underestimated the final state. The side-view perspective limited comprehension of total extent.

Participant Preference

Accuracy preference:
VR: 100% (11 out of 11 participants)
2D: 0% (0 out of 11 participants)

Impact preference:
VR: 100% (11 out of 11 participants)
2D: 0% (0 out of 11 participants)

Interpretation: Unanimous preference across both metrics. Perfect alignment between objective performance (VR was measurably more accurate) and subjective preference (all participants chose VR).

Qualitative Feedback

VR Strengths:

• "I liked the birds eye view perspective"

• "The ability to move into the forest was particularly effective, as I felt that I could see the deforestation from multiple views"

• "Good visuals, easy to maneuver"

• "The real controllers in the scene with the overlay instructions was very helpful"

VR Navigation Issues:

• "Difficult for rotating around, fly was a bit weird" (multiple mentions)

• "Right joystick was very jumpy and disorienting" (3 participants)

• "Brown floating rectangle under trees was trippy when viewing things"

• "Couldn't figure out how to move back"

• "The honed in circle that would occur in the headset was a bit confusing" (tunneling vignette)

2D Visualization Issues:

• "2D version was only a side view, so I couldn't really tell anything about how things were changing"

• "It felt unfair to compare to 2D-viz given that it was sideways and not really how a typical 2D-viz would look like"

• "I think the trees could be tilted a little more, more of an isometric view like google maps"

Improvement Suggestions:

• Add percentage overlays (2 mentions)

• Fix right joystick rotation sensitivity (3 mentions)

• Enable backward movement

• Improve 2D comparison with vertical/top-down view (2 mentions)

• Zoom in/out feature

• Hold-down to scroll through years faster

• Automatic time-lapse slider

• More scale/broader area

• Carbon emissions correlation

• Contextual information beyond visualization

Impact & Significance

Research Contribution

This project provides empirical evidence for VR's educational effectiveness in environmental data communication through three key findings:

Quantitative Validation: The 16.3% accuracy improvement is statistically significant (p < 0.05) and demonstrates measurable benefit over traditional 2D approaches.

Controlled Methodology: Within-subjects design controls for individual differences, while the identical question (2023 loss asked in both formats) provides the cleanest comparison with 43.5% VR advantage.

Unanimous Preference: 100% preference rate is exceptionally rare in user experience research, suggesting strong user demand for VR environmental visualization.

Educational Impact

Better Comprehension:

• Lower average error (4.09% vs 4.89%)

• More consistent performance across participants

• Superior capture of dramatic changes (2020 period: 17.2% error)

Stronger Emotional Connection:

• 100% chose VR for "impact"

• Participants overestimated severity (39% vs 36%) rather than underestimating like 2D users

• Multiple mentions of feeling "immersed" in the experience

Enhanced Engagement:

• Active exploration vs passive viewing

• Multi-perspective understanding from flying navigation

• Temporal navigation empowerment through controller input

Real-World Applications

This approach can be applied to:

Climate Education: Sea level rise visualization (walk through flooded cities), glacier retreat over time, temperature change mapping in 3D geographic space

Conservation Communication: Wildlife habitat loss, coral reef bleaching progression, urban sprawl impact visualization

Policy & Advocacy: Evidence-based communication to policymakers, public awareness campaigns, fundraising for environmental organizations

Scientific Research: Data exploration for researchers, pattern discovery in spatial-temporal datasets, hypothesis generation through immersive analysis

Lessons Learned

Technical Lessons

Test Deployment Pipeline Early: Waiting until Week 5 to attempt Quest deployment nearly derailed the project with Mac SDK issues. Deploy a "Hello World" build to target platform in Week 1.

Mobile Shader Complexity Matters: URP/Lit shaders timeout on Quest. Mobile/Unlit reduced build time 85%. Profile early and optimize shaders for mobile VR from the start.

Input System Conflicts Are Subtle: Mixing old and new Unity Input System caused errors. Commit to new Input System entirely and migrate completely.

Camera Lock Breaks VR: Custom scripts that override camera position will break XR Origin movement. Understand XR systems thoroughly before adding custom movement logic.

Design Lessons

VR Navigation Is Critical: Despite superior accuracy, navigation issues frustrated users. The project could have performed even better with polished movement UX. Invest heavily in movement comfort and intuitiveness.

Fair Comparison Requires Fair 2D: The 2D side-view was suboptimal design. Participants explicitly noted the unfair comparison. Use best-possible 2D visualization for valid research comparison.

Subtle Changes Need Help: Both VR and 2D struggled with small differences (2005: 1.92% loss). Percentage overlays would improve both formats. Don't rely solely on visual perception for precise values.

Dramatic Changes Shine in VR: The 2020 period (28.64% loss) had best VR performance because immersion amplifies dramatic changes. VR excels at making magnitude tangible.

Research Lessons

Within-Subjects Design Powerful: Controls for individual differences but introduces potential order effects. Use counterbalancing if feasible (half VR first, half 2D first).

Identical Questions Are Gold: The 2023 question asked in both formats provided the cleanest comparison (43.5% improvement) with no confounding variables. Always include at least one identical question across conditions.

Qualitative Feedback Essential: Open-ended responses explained the quantitative results and revealed unexpected issues like tunneling vignette disorientation. Always collect qualitative feedback alongside metrics.

Preference and Performance Alignment: When subjective preference matches objective performance (both favoring VR), the evidence is extremely strong. This alignment is rare in UX research.

Future Work

Immediate Improvements

Navigation Refinement:

• Fix right joystick sensitivity for smooth rotation

• Add backward movement capability

• Reduce movement speed from 50 to 10-15 units/sec

• Disable or reduce tunneling vignette effect

• Add in-game speed adjustment slider

Visual Enhancements:

• Add optional percentage overlays for precision

• Implement smooth time-lapse animation between years

• Add minimap showing user position

• Include ground texture beyond flat plane

2D Visualization Redesign:

• Switch from side-view to top-down or isometric

• Enable zoom and pan

• Match VR's year navigation UX

• Re-run comparison study with improved 2D

Long-Term Extensions

Expanded Geographic Coverage:

• Add 5-10 additional Rondônia tiles

• Create seamless larger forest environment

• Enable teleportation between regions

• Comparative view of multiple regions side-by-side

Enhanced Data Integration:

• Import tree cover gain data (reforestation)

• Add biodiversity impact indicators

• Integrate carbon emissions data

• Show correlation with deforestation drivers

Educational Features:

• Guided tours with narration

• Quiz mode with scoring

• Before/after comparison slider

• Save and share user perspectives

Research Extensions:

• Larger sample size (n=50+)

• Counterbalanced within-subjects design

• Delayed retention testing

• Behavioral outcome measurement

• Cross-age group comparison studies

Downloads & Resources

Quest APK:

• File: AmazonVR.apk

• Size: ~150 MB

• Requirements: Meta Quest 1/2/3

• Installation: Sideload via SideQuest or ADB

Source Code:

• Unity Version: 6.3 LTS

• Includes: All scripts, prefabs, and data processing code

• License: MIT

Data Processing Pipeline:

• Python scripts for GeoTIFF to PNG conversion

• Requires: rasterio, numpy, Pillow

• Input: Hansen GFC GeoTIFF

• Output: Unity-ready PNG files

User Study Materials:

• Google Forms template

• Excel analysis spreadsheet with formulas

• Study protocol documentation

References

1. Hansen, M. C., et al. (2013). High-resolution global maps of 21st-century forest cover change. Science, 342(6160), 850-853.

2. Global Forest Watch. (2024). Hansen Global Forest Change v1.11 (2000-2023). University of Maryland, NASA, USGS.

3. Unity Technologies. (2024). XR Interaction Toolkit Documentation.