Multispectral Dataset Preparation for Anomaly Detection

Goal:
Prepare clean multispectral datasets (WV3) for Isolation Forest anomaly detection of illegal waste, by removing clouds, water, and bright urban zones from all scenes.
Freelancer must process 4 multispectral scenes:
2022‑06‑13
2023‑06‑26
2024‑06‑29
2025‑08‑25 (August scene — cleanest, used for testing)
All scenes are already converted to points (x, y, band_1 … band_8).
Detect and remove cloud pixels
Use any reliable method (heuristic or ML), such as:
brightness threshold (top 1–2% in Green band)
NDVI < 0.05
NIR ratio (NIR2/NIR1) for cloud/ice separation
or simply manual polygon selection (if detection fails)
Output: one column "cloud" (True/False)
Remove water pixels
Use NDWI > 0.1 (or similar threshold). Output: "water" pixel flag

Remove bright urban zones
Typically detected using:
very high reflectance in band_3 or band_5
OR high BSI index
Output: "urban" pixel flag
Produce a clean dataset
Remove all:
invalid pixels (NaN, negative, all-zero bands)
cloud pixels
cloud shadows
water
bright urban zones
extreme outliers (per band: below 1% or above 99%)
Output: "CLEAN" dataset ready for Isolation Forest.