Daksh17440/satellite_images_color_clustered

🛰️ Mosaic Library — LAB-Clustered Satellite Tile Dataset

A curated collection of satellite image tiles harvested from ArcGIS World Imagery and organized into perceptually-uniform LAB color-space clusters. Each tile is a 256×256 PNG snapshot of Earth's surface at zoom level 15 (~4.8 m/pixel), deduplicated and routed into one of 5,000 color buckets using K-Means centroids in CIE LAB space. The dataset currently contains tiles covering 786 out of 5,000 color clusters, with up to 100 unique images per bucket.

📐 Dataset Structure

mosaic_library/
├── tiles.db               # SQLite index of all tiles (id, bucket, path, phash, LAB values)
├── centroids.npy          # NumPy array of 5,000 LAB cluster centroids (shape: 5000 × 3)
├── temp/                  # Ephemeral download scratch space (empty at rest)
├── bucket_0000/           # Up to 100 PNG tiles whose mean LAB color maps to centroid 0
├── bucket_0001/
│   └── <uuid>.png
...
└── bucket_4999/

Each bucket folder holds images whose dominant color, expressed in CIE LAB space, is nearest to that bucket's centroid. Folder names are zero-padded to four digits (bucket_0786).

SQLite Schema (tiles.db)

The database provides fast metadata lookups without loading any images.

Column definitions:

| Column | Type | Description |
|-------|-----|------------|
| `id` | TEXT (PK) | UUID v4 — matches the PNG filename |
| `bucket_id` | INTEGER | Cluster index, 0 to 4999 |
| `filepath` | TEXT | Relative path to the PNG file |
| `phash` | TEXT | 64-bit perceptual hash (hash_size=8) |
| `mean_l` | REAL | Mean L channel of the tile in CIE LAB |
| `mean_a` | REAL | Mean a channel |
| `mean_b` | REAL | Mean b channel |

Indexes: idx_bucket on bucket_id; idx_phash on (bucket_id, phash) for fast duplicate checks.

---

Quick queries:

import sqlite3
 
conn = sqlite3.connect("mosaic_library/tiles.db")
 
# Total tile count
conn.execute("SELECT COUNT(*) FROM tiles").fetchone()
 
# Buckets fully populated at 100 images
conn.execute(
    "SELECT bucket_id, COUNT(*) as n FROM tiles GROUP BY bucket_id HAVING n >= 100"
).fetchall()
 
# All tiles in bucket 42
conn.execute("SELECT * FROM tiles WHERE bucket_id = 42").fetchall()

---

Color Clustering Design

Why CIE LAB?

RGB is not perceptually uniform — equal numeric distances do not correspond to equal perceptual differences. CIE LAB is specifically designed so that Euclidean distance in LAB space closely matches human color perception. This makes the clusters visually meaningful: each bucket represents a genuinely distinct visual texture-color region of Earth's surface.

---

How Clusters Were Generated:

import numpy as np
from skimage.color import rgb2lab
from sklearn.cluster import MiniBatchKMeans
 
random_rgb = np.random.randint(0, 255, (500000, 1, 3), dtype=np.uint8)
random_lab = rgb2lab(random_rgb).reshape(-1, 3)
 
kmeans = MiniBatchKMeans(n_clusters=5000, batch_size=10000, n_init=3)
kmeans.fit(random_lab)
 
centroids = kmeans.cluster_centers_
np.save("mosaic_library/centroids.npy", centroids)

---

At inference time, a KD-Tree (scipy.spatial.cKDTree) provides O(log n) nearest-centroid lookup for each incoming tile.

---

Tile Assignment Pipeline

Each downloaded tile goes through this pipeline before being saved:

1. Resize to 64x64 pixels for fast color computation.
2. Compute mean RGB across all pixels and convert to LAB.
3. Query the KD-Tree to find the nearest centroid and assign `bucket_id`.
4. If the bucket already holds 100 images, discard the tile.
5. Compute perceptual hash (`imagehash.phash`, hash_size=8) and discard if a matching hash exists in the same bucket.
6. Atomically move the file into `bucket_XXXX/` and insert a row into `tiles.db`.

---

Geographic Coverage

Tiles were sampled across two harvesting phases with different biome weightings.

Phase 1 — General Biomes

| Biome | lon_min | lat_min | lon_max | lat_max | Weight |
|------|--------|--------|--------|--------|------|
| Desert (Sahara) | -10 | 15 | 30 | 30 | 15% |
| Forest (Amazon/Congo) | -70 | -15 | -50 | 5 | 15% |
| Urban (Europe/US) | -5 | 40 | 20 | 55 | 20% |
| Arctic (Greenland) | -50 | 65 | -30 | 80 | 10% |
| Grassland (US Midwest) | -105 | 35 | -90 | 45 | 15% |
| Shallow Water (Caribbean) | -80 | 15 | -65 | 25 | 10% |
| Deep Ocean (Pacific) | -150 | -20 | -110 | 20 | 5% |
| Agriculture (India) | 73 | 20 | 80 | 30 | 10% |

Phase 2 — Spectral Anomalies (Rare Colors)

Phase 2 targeted visually extreme locations to populate underrepresented LAB clusters:

| Location | Visual Signature | Weight |
|---------|-----------------|------|
| Australian Outback | Rust red | 20 |
| Salar de Uyuni, Bolivia | Bright white salt flat | 20 |
| Lake Natron, Tanzania | Pink/red alkaline water | 15 |
| Dallol, Ethiopia | Neon yellow-green hydrothermal | 15 |
| Yellowstone Hot Springs | Cyan and orange rings | 15 |
| Bahamas Shallows | Aqua/teal shallow water | 15 |
| Lava Fields, Hawaii | Pitch black basalt | 10 |
| Atacama Desert, Chile | Ochre/orange arid | 10 |
| Dense Urban Tokyo | Concrete grays | 10 |
| Greenland Glaciers | Ice blue/white | 10 |

---

🔧 Data Collection Parameters

Parameter	Value
Tile source	ArcGIS World Imagery (MapServer/tile/{z}/{y}/{x})
Zoom level	15 (~4.8 m/pixel ground resolution)
Tile size	256 × 256 px (ArcGIS default)
Format	PNG (RGB)
Total color buckets	5,000
Max tiles per bucket	100
Buckets populated (current)	786 / 5,000
Deduplication method	Perceptual hash (pHash, 64-bit) per bucket
Download concurrency	20 threads
Batch size per iteration	100 tiles

---

🐍 Loading the Dataset

Load all tile metadata

import sqlite3
import numpy as np
 
conn = sqlite3.connect("mosaic_library/tiles.db")
rows = conn.execute("SELECT id, bucket_id, filepath, mean_l, mean_a, mean_b FROM tiles").fetchall()
print(f"Total tiles: {len(rows)}")

Load tiles from a specific bucket as NumPy arrays

from PIL import Image
import numpy as np
import sqlite3, os
 
def load_bucket(base_dir, bucket_id, db_path="mosaic_library/tiles.db"):
    conn = sqlite3.connect(db_path)
    rows = conn.execute(
        "SELECT filepath FROM tiles WHERE bucket_id = ?", (bucket_id,)
    ).fetchall()
    images = []
    for (fp,) in rows:
        img = Image.open(os.path.join(base_dir, fp)).convert("RGB")
        images.append(np.array(img))
    return np.stack(images)   # shape: (N, 256, 256, 3)
 
imgs = load_bucket(".", bucket_id=42)
print(imgs.shape)

Query by LAB color proximity

from scipy.spatial import cKDTree
import numpy as np
 
centroids = np.load("mosaic_library/centroids.npy")
tree = cKDTree(centroids)
 
# Find the bucket closest to a target LAB color
target_lab = np.array([60.0, -20.0, 30.0])  # e.g. greenish mid-tone
dist, bucket_id = tree.query(target_lab)
print(f"Nearest bucket: {bucket_id}  (distance: {dist:.2f})")

---

⚠️ Limitations & Notes

• 786 of 5,000 buckets are currently populated. Many LAB colors are extremely rare in nature (e.g. pure magenta terrain) and may never fill without targeted synthetic augmentation.
• Tiles derive from ArcGIS World Imagery (ESRI). Usage is subject to ESRI's terms of service. This dataset is intended for non-commercial research purposes.
• Ocean and cloud tiles may appear in general biome samples; they are not explicitly filtered beyond color-bucket routing.
• Perceptual hash deduplication is per-bucket only — the same geographic location could theoretically appear in two different buckets if its color shifts (e.g. seasonal change), though this is rare at zoom 15.
• The centroids.npy file is required to correctly interpret bucket assignments. Do not regenerate it, as this would invalidate all existing bucket assignments.