---

🌳 Quantifying the "Invisible Forest"

3D Biometric Gap Analysis & Urban Digital Twin of Hong Kong's High-Density Canopy

📌 Executive Summary

In the hyper-dense urban fabric of Hong Kong, a significant portion of the vegetated landscape exists as an "Invisible Forest"—trees that provide critical ecosystem services and pose potential structural risks but remain omitted from formal municipal registers.

This project establishes a Deterministic Digital Twin to identify the biometric gap between the Physical Reality (detected via 2020 Airborne LiDAR) and the Administrative Record (Municipal Tree Registers). By focusing strictly on Parks, Roads, and Slopes managed by LCSD, HyD, ArchSD, CEDD, and DSD, this study seeks to isolate the "delta" of unmanaged trees.

---

🛠️ Technical Architecture

1. Spatial Foundation & Coordinate System

• CRS: Hong Kong 1980 Grid (EPSG:2326)
• LiDAR Baseline: CEDD 2020 Territory-wide Survey (.laz)
• Land Use Mask: PlanD 2020 Raster Grids (LUM_end2020.tif)

2. The 4-Phase Workflow

Stage	Focus	Primary Tools
Stage 1	Jurisdictional Masking	sf, raster, lidR
Stage 2	2D AI Segmentation	SAM, YOLOv11-seg
Stage 3	3D Biometric Extraction	lidR (Alpha-shapes, LMF)
Stage 4	Administrative Audit	Geospatial Left-Join

---

🚀 Key Methodology

🛡️ Stage 1: The Liability Mask (Cookie-Cutter Logic)

To ensure the audit is administratively relevant, we use a multi-layered mask to exclude private land:

• Included: Parks (LUM 41/42), Roads (LUM 51), Infrastructure (LUM 61), and SIMAR Slopes.
• Excluded: Private Residential (LUM 1) and Public Housing (LUM 2).
• Filter Logic: $Liability Zone = (Public Infrastructure \cup Slopes) \setminus Private Estates$

🤖 Stage 2: Instance Segmentation (SAM + YOLO)

We use a SAM-to-YOLOv11-seg pipeline to solve the "Cluster Problem" of interlocking broadleaf canopies (e.g., Ficus microcarpa). This provides pixel-perfect polygons instead of simple bounding boxes, allowing for cleaner 3D LiDAR extraction.

📊 Stage 3: 3D Biometrics

Using the AI-generated polygons, we "cookie-cut" the 1.7GB LiDAR tiles.

• Height Threshold: $H_{min} > 5.0m$ (Deletes grass, shrubs, and small saplings).
• Metrics: Maximum Height ($H_{max}$), 3D Crown Volume ($V_{crown}$), and Vertical Distribution Ratio ($VDR$).

---

💻 Code Snippets (Laptop-Safe LiDAR Processing)

# Setup the out-of-core processing engine for large files
library(lidR)

# Initialize the catalog virtually (16GB RAM Safe)
ctg <- readLAScatalog("./data/lidar_tiles/")

# Configure the engine
opt_chunk_size(ctg)   <- 250   # 250m squares
opt_chunk_buffer(ctg) <- 30    # 30m overlap for whole crowns
opt_select(ctg)       <- "xyz" # Load only coordinates
opt_filter(ctg)       <- "-keep_class 2 3 4 5" # Ground + Veg only

# Normalize Height (Find the true height above ground)
normalized_ctg <- normalize_height(ctg, tin())

---

📂 Repository Structure

├── 01_masking/             # Raster-to-Vector jurisdictional masking
├── 02_segmentation/        # SAM & YOLOv11-seg training scripts
├── 03_biometrics/          # lidR scripts for volumetric extraction
├── 04_audit/               # Statistical gap analysis (Rmd)
├── assets/                 # Legends, CRS definitions, and documentation
└── README.md

📋 Requirements

• R 4.0+: lidR, sf, raster, dplyr
• Python 3.9+: ultralytics (YOLOv11), segment-anything, geopandas
• Hardware: 16GB RAM minimum (32GB recommended for LiDAR normalization).

---

📝 Acknowledgements

• Data Providers: Hong Kong CSDI Portal & CEDD Spatial Data Portal.
• Technical Frameworks: Based on the lidR package (Silva et al.) and the YOLO computer vision framework.

---