Exploratory Data Analysis: Rock Glacier Dataset
1. Basic Dataset Information
| Property | Value | Description |
|---|---|---|
| Total Features | 626 |
The total number of rock glacier polygons in the dataset. |
| Geometry Type | Polygon |
The geometric shape format of the data. |
| CRS | EPSG:4326 |
Coordinate Reference System (Standard Latitude/Longitude, WGS 84). |
| Bounding Box | [88.0747, 27.3212] to [88.9281, 28.0849] |
The geographic extent (Longitude / Latitude). This falls roughly in the Sikkim Himalayas region. |
2. Attribute Table Analysis
The shapefile contains 20 columns, many of which are empty or artifacts from a previous format conversion (likely KML/KMZ to Shapefile).
Useful Columns
These columns contain data that may be relevant for identifying or tracking the source of the polygons:
| Column Name | Data Type | Description & Preview |
|---|---|---|
Name |
String | IDs or names (e.g., "1", "2", "NULL01"). |
id |
String | A secondary ID field present in some rows. |
layer |
String | Source layer name (e.g., finl_sikm_shp). |
path |
String | The original file path of the source data. |
geometry |
Geometry | The actual polygon boundary coordinates. (Crucial for deep learning masks) |
Unused / Empty Columns
The following columns are mostly filled with None, NaN, or default values (1, 0, -1) and can be safely ignored or dropped during the data preprocessing phase:
- Empty Data:
descriptio,timestamp,begin,end,altitudeMo,shp_05,descript_1,altitude_1,RG_AJ,lohanak - KML Artifacts:
tessellate,extrude,visibility,drawOrder,icon
3. Data Preview (First 5 Rows)
Below is a cleaned snippet of the first 5 entries showing only the most relevant columns, instead of the messy raw pandas output:
| Index (Hidden ID) | Name | layer | path | id | geometry |
|---|---|---|---|---|---|
| 0 | 1 | NaN | NaN | NaN | POLYGON Z ((88.0865... |
| 1 | 2 | NaN | NaN | NaN | POLYGON Z ((88.0864... |
| 2 | 3 | NaN | NaN | NaN | POLYGON Z ((88.0863... |
| 3 | 3 | NaN | NaN | NaN | POLYGON Z ((88.0863... |
| 4 | 4 | NaN | NaN | NaN | POLYGON Z ((88.0862... |
4. Recommendations for Deep Learning Pipeline
- Geometry Rasterization: For your deep learning models (U-Net, SegFormer), you only need the
geometrycolumn. These polygons will be rasterized to create binary masks where1 = Rock Glacierand0 = Background. - Coordinate Projection: Since the dataset is in degrees (
EPSG:4326), ensure that your input rasters (Sentinel-2, DEM) match this CRS. If you need to extract patches of a specific size in meters (e.g., 256x256 pixels at 10m/pixel), you should project this shapefile to a local UTM zone (in meters) before patching. - Data Cleaning: You can completely drop the 15+ unused KML columns when loading this shapefile with
geopandasto save memory and keep the pipeline clean.