DrivenData-OpenAI-Caribbean-Challenge

Driven Data Competition - AI Caribbean Challenge: Mapping Disaster Risk from Aerial Imagery

LB score: 1.3307 | LB Position: 100 amongst 1425 competitors

Note: clean code will be uploaded soon!!

In this challenge, your goal is to use provided aerial imagery (GeoTiff) to classify the roof material of identified buildings in St. Lucia, Guatemala, and Colombia.

Data

• Metadata: CSV with metadata linking images (GeoTiffs) with corresponding train/test label files (GeoJSONs)

• Train Labels: The training labels in CSV format where each row contains a unique building ID followed by roof material (one hot encoded columns)

• Submission Format: The submission format

• STAC: A SpatioTemporal Asset Catalog of the imagery and labels

ML Workflow

1. Download GeoTIFF + GeoJSON from STAC

2. Summarize train data info (tiff_path, geo_path per roof_id etc)

3. Clip image tiles from Geotiffs based on GeoJSON data (transform ccoords system whenever it is required) x 7 regions

4. Reshape images with various sizes to (IMG_SIZE, IMG_SIZE, 3) for training --> needs extra care !!

5. Form X, Y, i.e. features (pixels) and labels

6. Save X, Y, to pickle for future use..

7. Feature Engineering - image transformations (optional)

8. Build DL model

9. Train and evaluate model

10. visualize and assess predictions