TASK_02_Feature_Extraction

Assignee: Google Earth Engine Specialist
Role: Feature Extraction & Stack Generation

1. Project Overview

We are building a Rock Glacier Segmentation Dataset for Himalayan rock glacier mapping. The segmentation labels have already been generated from manually delineated rock glacier polygons.

Current Project Structure:

RockGlacier_Project/
│
├── EDA/
│   ├── individual_polygons/
│   │   ├── polygon_0.geojson
│   │   ├── polygon_1.geojson
│   │   └── ...
│   └── individual_plots/
│
├── masks/
│   ├── mask_0000.tif
│   ├── mask_0001.tif
│   └── ...
│
├── features/
└── dataset/

Polygons: Located at EDA/individual_polygons/ (e.g., polygon_0.geojson)
Masks: Already generated and located at masks/ (e.g., mask_0000.tif)
Mask Values: 0 = Background, 1 = Rock Glacier

2. Objective

Generate georeferenced multi-band feature stacks from Google Earth Engine for every rock glacier polygon. The generated features must align perfectly with their corresponding masks.

Important Notes This is a semantic segmentation project.

Do NOT create polygon-level statistics.

Do NOT calculate mean NDVI, mean slope, mean elevation, etc.

Do NOT generate CSV tables.

We explicitly need pixel-level feature images (rasters).

3. Data Sources & Required Features

Use Google Earth Engine to extract the following:

Primary Imagery: Sentinel-2 Surface Reflectance
Terrain: Copernicus DEM (preferred)

Required Bands/Features:

Sentinel-2 Bands:

Blue (B2)
Green (B3)
Red (B4)
NIR (B8)
SWIR1 (B11)
SWIR2 (B12)

Spectral Indices:

NDVI
NDWI
NDSI

Terrain Features:

Elevation
Slope
Aspect

If feasible, also include:

Land Surface Temperature (LST)
Surface Velocity

4. Technical Specifications

Spatial Resolution

All outputs must use exactly 10 meter spatial resolution.

Reason: Sentinel-2 native resolution is 10m. The masks and features must have absolutely identical spatial resolutions to prevent dimension mismatch during U-Net training.

Extraction Strategy

For each polygon, perform the following sequence:

Load polygon_i.geojson
Compute bounding box
Add 200 meter padding around the bounding box. (Note: This padding allows future segmentation models to see surrounding terrain context, not just the glacier).
Download all required features for this bounds area.
Clip features to the padded region.
Resample all layers to exactly 10m resolution.
Stack all features into a single multi-band GeoTIFF.

Output Location

Save all generated feature stacks inside the features/ directory:

features/
├── feature_0000.tif
├── feature_0001.tif
├── feature_0002.tif
└── ...

5. Band Order Specification

Your final GeoTIFF stacks must strictly follow this band ordering:

Band 1 = Blue
Band 2 = Green
Band 3 = Red
Band 4 = NIR
Band 5 = SWIR1
Band 6 = SWIR2
Band 7 = NDVI
Band 8 = NDWI
Band 9 = NDSI
Band 10 = Elevation
Band 11 = Slope
Band 12 = Aspect

Optional (if acquired):

Band 13 = LST
Band 14 = Velocity

6. Alignment Validation

You must physically verify that the features perfectly overlap the masks:

feature_0000.tif aligns perfectly with mask_0000.tif
feature_0001.tif aligns perfectly with mask_0001.tif

Checklist for alignment:

Matching CRS (Coordinate Reference System)
Matching Resolution (10m)
Matching Bounding Box Coordinates
Exact Pixel-to-Pixel alignment (dimensions must be identical)

7. Deliverables & Current Scope

What you must deliver:

Google Earth Engine extraction script.
Feature stack generation pipeline.
Validation script (to verify alignment checklist above).
Documentation explaining each generated feature.
Feature extraction report.

Strict Scope Limitations

Only perform feature extraction and feature stack generation.

Do NOT create train/validation/test splits.

Do NOT generate patches.

Do NOT train any machine learning model.

The sole goal of this task is to produce correctly aligned multi-band feature stacks for all rock glacier masks.